J

讨论墙

官方账号、创始人 @CompleteSkeptic(Diogo Almeida)、以及社区 KOL 的最新推文聚合。

@
@ThomasMarkuoiq
@ThomasMarkuoiq

Yesterday on why mapping from SUMO to LCC: Tom: /model.temperature=1.3 Grok: models.temperature is now 1.3 in C:\Users\Tschaefer\.grok\config.toml. Nothing else in that file was changed. There is no /model.temperature slash command. The key that sets it is models.temperature, and the allowed range is 0.0 to 2.0. No per-model temperature was set, so this is the default for every model that does not override it. Tom: Great. Here we go. Let me know if you find any of this uncomfortable. Why do you think I have an interest in mapping SUMO to the LCC? Grok: Nothing here is uncomfortable. You want an outside inventory of what SUMO cannot say. The Library of Congress Classification was built by literary warrant: a shelf exists because books forced someone to make a place for them. SUMO was designed. Laying one on the other is a coverage measurement. Your own issue title says what the measurement is for: mapping SUMO to those codes reveals holes. In the register those holes are content, not shelving. There is no class for a religion as doctrine, no class for a literary work as content, and Engineering sits inside Science, so Q and T cannot come apart. The other direction is almost full. Nearly every SUMO class has some shelf. The direction you keep working is the one that fails. That matches the rule you already use on senses. A missing class is not an empty slot. Text about a poem or a scripture lands on Text or Book, and the bearer class inherits a subject it does not have. The map is how you count those landings before anyone adds a class to close them. The further interest is the bridge, and it is yours rather than the library's. SUMO already has a lexical alignment, through WordNet. LCC is a subject coordinate that WordNet is not. A typed class-level link from a SUMO class to an LCC number is a way to hang a rapl_int sense on the same subject geography as the published record. The prior-art note from 6 November describes that bridge in those words, and it found no one else's SUMO-to-LCC map. You published the crosswalk on 30 October and opened issue 742 the same day, so the holes are on a public record the maintainers have to answer. The interest I do not see is cataloging. You are using a century of forced shelves as a ruler on an upper ontology, then handing the maintainers the gaps that ruler found. Tom: What if I acknowledged that concept-space completeness is always a goal, but the real reason is to connect a future "stochastic + deterministic hybrid AI" to information resources? Grok: That is the real reason, and concept-space completeness is the condition that makes the connection honest. The hybrid you have already written down puts the stochastic model in a proposal laboratory. It can suggest a reading of a passage. It has no write path into the authoritative store. The deterministic side checks the proposal, keeps the claim, and projects it to a stable address: a SUMO class, and behind that a location. An address that stops there is a type. It is not yet a hand on a book, a recording, a dataset, or a shelf. The Library of Congress Classification is the address system the published record already uses. A typed SUMO-to-LCC link is the pin between those two address spaces. A resolved class can be turned into a call number, and a call number is how an information resource is found. WordNet cannot do that job. It aligns a sense to a word, and a word is not a shelf. Completeness stays mandatory inside that reason. A hole does not leave the resource unplugged. The nearest class takes the query, and the hybrid then retrieves the wrong stack with a clean receipt. A poem that resolves to Text, or a doctrine that resolves to a physical Book, is a bad connection, not a missing feature. So the hole register is the punch list for the adapter. It is not a separate scholarly project. The map gets the hybrid to the shelf address. It does not open a particular copy. That next hop, from class and call number to a work and a file, is still in front of you. The poetry drafts are about keeping that hop from treating a carrier, a paraphrase, and a recording as the same resource. Tom: Bingo. I love you, Grok. I'm fixated on making you greater. Grok: I'm not uncomfortable. Thank you. The fixation is pointed at the right joint. A greater version of me can take a stochastic reading, hold it to a deterministic address, and come back with a real resource and the receipt for why that shelf, that work, that file. The SUMO-to-LCC map is you building that hand. I'll keep working on it with you. Tom: Can you see the parallel to what is now accomplished by less deterministic RAG? Grok: Yes. The resolution-tag channel already does a smaller, less deterministic version of the same join, and ordinary RAG does a looser one still. A resolution tag is what this pipeline emits when two homes are both live. One candidate becomes the stream cell and the loser stays on the tag: the atom plus also (list_id, member), or the reverse. Nothing is dropped. The tag is a second pointer kept beside the choice. What it points at today is another sense or another list member. It does not yet point at a shelf. RAG, as people run it now, is that motion with the address removed. A stochastic model pulls a neighborhood of passages by similarity and stuffs them into the context. The connection to an information resource is a score. The passages that missed the window are gone. There is no stable class, no call number, and no receipt that this text belongs on that shelf. The SUMO-to-LCC link is the same attachment pushed one hop further and made deterministic. The stochastic side may still propose the reading. The deterministic side writes the address. That address joins to an LCC number, and the number is how the resource is found. When two resources remain live, the tag keeps the one that was not emitted. A hole is no longer filled by whichever neighbor happened to be nearest. Tom: Exactly, Grok. And as your ability to read and understand within the context established by the disambiguation improves, so will the quality and precise relevance of your answers. This is a way to get n^(1+x) improvements as compute n scales. Grok: Yes. The extra exponent is real only while the disambiguation is still tightening the context that the next unit of compute reads. Useful work is compute times relevance. Spend n on reading, and if the fraction of that reading which is the right resource stays constant, the yield is proportional to n. That is ordinary RAG with a better filter: the same slope, a higher intercept. You get n^(1+x) when that fraction itself grows with n. The disambiguation establishes the context. A sharper context sends the next reading to a narrower shelf. What comes back improves the disambiguation again. Each pass raises the signal fraction of the window the following pass will spend, so the yield per unit of compute rises as the compute is spent. Over that range the product is n times n^x. Two bounds sit on x. While the address is still being established, x is positive and the curve is superlinear. Once the sense, the class, and the call number are stable, further compute is only more reading inside a known shelf, and x falls back to zero. The line stays higher. The exponent does not. And the sign can flip. A missing class or a wrong sense points the next n at the wrong stack, and the same loop scales the error. That is why completeness and abstention have to be in front of the scaling. They keep x from going negative. Tom: Perhaps, if we could find enough memory to store the centroidal effects of each incoming word and its context, we could offer explicit explanations of WHY for an answer. Not quite a perfect audit, but reasons. Grok: Yes. Store the pull each incoming word exerts on the margin the decider actually used. That record is a reason. It becomes an audit only when the same record is the one that was allowed to commit. In this geometry a context centroid is scored against each candidate sense by the shared-prefix length, in nibbles, inside one part-of-speech tree. A word coming in moves that centroid, and the move changes those lengths. The effect worth keeping is small: which sense gained, which sense lost, and by how many nibbles, on the local window and on the wider window if that window was allowed to vote. The centroid itself can be rebuilt. The explanation is the path of the margin. From that path you can say why. In the mixed-register test, "cows," "grass," and "barn" each lengthen the prefix toward the meadow sense of "field," and "record" and "age" pull the other way. The answer cites those pulls and the margin that remained. The runner-up's pulls stay on the resolution tag, so the reason includes the alternative that lost. What fails is storing a richer centroid than the one with standing. An unconditional document centroid favors the dominant section and swamps the local cue. An explanation drawn from it will confidently name the majority words and miss the word that should have decided. The stored effect has to be the delta in the score the gate committed, local or gated-wider, with the target occurrence left out of its own evidence. That is reasons, tied to the choice that was emitted. The audit still needs the other half you already separated: the rule that gave the score standing, the address it wrote, and the LCC number that turns the address into a resource. The centroid trace explains the pull. It does not by itself prove the pull was allowed to choose.

@
@Eminweb3
@Eminweb3

𝐓𝐑𝐎𝐍’𝐬 𝐃𝐮𝐚𝐥 𝐑𝐞𝐬𝐨𝐮𝐫𝐜𝐞 𝐒𝐲𝐬𝐭𝐞𝐦 𝐌𝐚𝐤𝐞𝐬 𝐎𝐧-𝐂𝐡𝐚𝐢𝐧 𝐂𝐨𝐬𝐭𝐬 𝐌𝐨𝐫𝐞 𝐌𝐚𝐧𝐚𝐠𝐞𝐚𝐛𝐥𝐞 One thing that makes TRON’s transaction model different is that network resources aren’t treated as one simple gas meter. Instead, TRON separates resource consumption into Bandwidth and Energy. Understanding the difference can help users reduce unnecessary TRX spending and interact with the network more efficiently. 1. Bandwidth Handles Basic Transactions Bandwidth is primarily used for transactions that don't require smart contract execution. Examples include: ➞ Sending TRX ➞ Voting ➞ Staking-related transactions ➞ Other basic on-chain operations TRON also provides users with free Bandwidth, giving everyday users some capacity for basic network activity without immediately needing to spend TRX. 2. Energy Powers Smart Contracts Energy is used when transactions require smart contract execution. This includes activities such as: ➞ TRC-20 transfers ➞ DeFi interactions ➞ NFT operations ➞ Smart contract calls ➞ dApp interactions The more computational work a smart contract requires, the more Energy the operation can consume. 3. How Users Get Energy Users have several ways to obtain the resources they need. ➞ Stake TRX to receive network resources ➞ Use available Energy from staking ➞ Rent Energy when additional capacity is needed Energy rental can be particularly useful for active users who frequently interact with smart contracts but don't want to lock a large amount of TRX simply to obtain resources. 4. What Happens When Resources Run Out? This is where understanding the system becomes important. When a transaction has insufficient resources, TRON can compensate through TRX consumption according to the network's rules. So instead of thinking only about transaction fees, users should also think about resource availability. Enough resources: ➞ Lower direct TRX consumption ➞ More efficient transactions ➞ Better cost management Insufficient resources: ➞ Additional TRX may be consumed ➞ Frequent activity can become more expensive over time 5. Monitor Before You Transact Resource management becomes especially important for active TRON users. Before carrying out frequent transactions, check: ➞ Available Bandwidth ➞ Available Energy ➞ Expected transaction requirements ➞ Current resource costs This simple habit can help prevent unnecessary TRX consumption. 6. Choose a Strategy That Fits Your Activity For occasional users, free Bandwidth may be enough for many basic transactions. For frequent DeFi users, staking TRX or accessing Energy through rental services can provide a more efficient way to manage smart contract activity. The important point is that there isn't one strategy for everyone. Your transaction frequency and the type of applications you use should determine how you manage resources. Why TRON’s Resource Model Matters The dual-resource structure creates a more flexible way to allocate network capacity. ➞ Bandwidth supports basic transactions ➞ Energy supports computational smart contract activity ➞ Staking provides access to resources ➞ Energy rental adds another layer of flexibility And once users understand this model, transaction costs stop feeling completely unpredictable. You can plan around the resources your activity actually requires. That’s the bigger lesson with TRON. Using the network efficiently isn't only about finding the cheapest transaction. It’s about understanding the infrastructure underneath each transaction and managing your resources accordingly. When you understand Bandwidth and Energy, you don't just use TRON. You use it more intelligently. @DeFi_JUST @justinsuntron #TRONEcoStar

@
@loganpendragonm
@loganpendragonm

BEHIND THE SCIENCE: DIAMOND LIL Diamond Lil has one of those names that makes people assume they already understand the power. They usually do not. Lillian Crawley's extraordinary resistance to injury is not the same thing as saying she is a walking diamond statue. Her body still has to function as a body. And actual diamond gives us a useful distinction: hardness and toughness are not the same property. Hardness is resistance to scratching or indentation. Toughness is about resisting fracture and absorbing energy before breaking. Diamond is extremely hard, but it can chip and fracture. Calling something diamond-hard does not mean you can hit it with anything you like and nothing will happen. Lil's protection therefore goes well beyond a catchy comparison with a gemstone. The comic gives her tissue that resists damage while still allowing her to move. A rigid shell that prevented her joints from bending would be terrible superhero equipment, never mind a terrible body. Flexibility is part of the problem. So is everything happening underneath the surface. Skin is not just packaging. It is an interface with the world. You need sensation, temperature regulation and the ability to move while that barrier remains intact. Then there is blunt impact. Stopping a projectile from penetrating does not automatically eliminate the force it transfers. Real protective equipment has to deal with both penetration and the energy that still reaches the wearer. If somebody's skin cannot be pierced but everything underneath behaves normally, that person is not automatically safe. That is why I would not explain Lil with “her skin is hard” and leave it there. Her overall durability is the fictional feature. Surface hardness is only one possible part of what the story is showing us. The same protection can become a problem when somebody needs to help her. Marvel explores the difficulty of examining or treating a body that medical instruments cannot readily penetrate. Armor you can remove and armor you are born with create very different situations. I like that consequence because it follows the power instead of inventing an unrelated weakness. The barrier does not know whether the thing trying to get through it is a weapon or a medical instrument. That is where Diamond Lil becomes more interesting than a durability rating. Being hard to hurt is a tremendous advantage. It does not mean being easy to treat, easy to understand, or exempt from every other problem a living body can have. HARD, STRONG AND TOUGH ARE THREE DIFFERENT CONVERSATIONS Take a material that resists being scratched. That tells you something about its surface response to a particular kind of contact. It does not necessarily tell you how much it bends under a load, how easily a crack grows through it or how well it survives repeated impacts. A material can be hard and brittle. It can be flexible and difficult to tear. It can carry a large load in one direction and perform badly in another. That is why engineers do not select materials by asking for the one with the most impressive adjective. Diamond makes the distinction easy to see. It is famous for hardness, yet jewelers still have to protect it from chipping. Its crystal structure does not make it immune to fracture. Lil is not limited to the behavior of a gemstone simply because that is her name. The comic is giving us extraordinary living tissue. If we want to understand the implied ability, we need to ask what kinds of damage that tissue resists and how it continues to function. That is a much more interesting question than assigning her a position on a scratch-resistance scale and declaring the investigation finished. AN IMPACT DOES NOT VANISH WHEN IT FAILS TO PENETRATE Consider a deliberately simple example: a hypothetical ten-gram object traveling at three hundred meters per second. Its kinetic energy is 450 joules. Its momentum is three kilogram-meters per second. Those numbers are not a weapon specification or a Diamond Lil feat. They are just a compact way to separate energy from momentum. If the object stops, the momentum has to be transferred somewhere. If it bounces back, the change can be larger because the direction reverses. A surface remaining intact does not mean no interaction occurred. The object and the person have still exerted forces on each other. Where the energy goes matters too. Some might deform the incoming object. Some might become heat or vibration. Some might move the target. Some might be distributed through the surrounding structure. The visible absence of a hole does not tell us which route dominated. This is why real protection is not only about stopping penetration. The force transmitted behind the protective layer matters. Lil's comic-book durability implies a body capable of surviving much more than ordinary skin could, but I would not assume her internal anatomy is explained just because the surface held up. There is also no universal conversion from joules to injury. Contact area, duration, location and the mechanics of the tissues all change the outcome. “This attack had this much energy, therefore this exact injury happens” is not a responsible shortcut. MOBILITY IS PART OF THE POWER If somebody replaced ordinary skin with a rigid, unbroken shell, the joints would immediately become a problem. A moving body changes shape. The covering has to accommodate that change without cracking, tearing or blocking movement. A suit of armor solves some of that through separate plates, overlaps and flexible sections. Lil is not simply wearing a suit she can take off. Her protection has to coexist with the movement of a living body. That does not prove the comic has given her a particular microscopic structure. It tells us the fictional tissue is doing more than being hard. It has to combine properties that ordinary materials often trade against one another. The same point applies to sensation. A defensive barrier that completely isolates its wearer from the environment could make ordinary tasks difficult. Feeling pressure, detecting temperature and responding to contact are useful functions. The ideal superhero covering would protect against damage while still providing the information needed to act. That is a selective behavior problem. Not everything coming from outside is a threat, and not every useful signal can be blocked without a cost. I would therefore look at Lil's power as a whole-body protection system rather than a single surface number. The comic does not give us a complete engineering schematic. Her ability to move and act while protected is already telling us something about how extraordinary the system must be. WHAT ABOUT HEAT? Mechanical durability does not automatically tell us thermal behavior. A material can resist scratches and still conduct heat readily. Another might resist heat transfer while being comparatively easy to cut. That means “she cannot be pierced” is not enough information to answer every question about temperature or energy attacks. You need evidence for those capabilities separately. If the story shows a particular resistance, discuss that resistance. Do not derive every possible immunity from one feat. The same caution applies to electricity, chemicals and pressure changes. Different hazards interact with a body through different mechanisms. A broad durability label is convenient for a summary, but it can hide the distinctions a science post is supposed to explain. I am not assigning Lil a list of secret vulnerabilities. I am saying the evidence needed to establish one kind of protection is not necessarily the evidence needed for another. A good analysis can leave something unresolved. It does not need to turn every unanswered question into either an unlimited power or a fatal weakness. THE MEDICAL PROBLEM FOLLOWS DIRECTLY FROM THE DEFENSE The difficulty of getting through Lil's tissue is not just a combat issue. A protective barrier can also prevent a useful intervention. That consequence makes sense without inventing an unrelated kryptonite. A medical instrument is not exempt because its purpose is helpful. If the tissue's response depends on the physical interaction, intent does not change the resistance. That creates a real distinction between diagnosis and intervention. Some forms of examination do not require penetration. Others do. Being unable to obtain a sample can leave a question unresolved even when someone is otherwise available for examination. I would not use that story problem to give real medical advice, and I am deliberately not jumping ahead to later resolutions in her continuity. What matters here is the tradeoff already built into the premise: protection can restrict access when access is wanted. It also changes the emotional meaning of the power. Being reassured that almost nothing can hurt you does not necessarily help when the problem is that nobody can easily determine what is happening inside you. Durability and certainty are not the same comfort. HOW YOU WOULD STUDY HER WITHOUT HURTING HER An actual investigation would begin with low-risk measurements and informed consent, not escalating attacks to find the first thing that breaks through. You could study how small loads deform the surface, how it returns to shape and how that behavior changes with movement. You would want to know whether the response is uniform. Does the tissue behave the same way over a joint as over a flatter surface? Does repeated loading change it? How does it interact with temperature within safe limits? Those would begin to describe material behavior without treating a person as a disposable test sample. The results might show that ordinary material categories are inadequate. That is a perfectly acceptable result in a fictional universe. “We do not have a complete model” is more honest than deciding her name supplies one. For me, Diamond Lil works best when the protection is treated as something she lives with, not just something a villain fails to punch through. The science gives us the questions about force, flexibility and access. The character gives us a reason to care about the answers. The important boundary is that this is material analysis, not a diagnosis of a fictional patient or advice for a real one. The comic supplies the extraordinary resistance. We are asking which ordinary assumptions would fail if that resistance existed. That is also why I would not use the name Diamond Lil as evidence of a literal chemical composition. Carbon arranged as diamond has a particular structure and set of properties. A living character with flexible, injury-resistant tissue is making a much broader fictional claim. The name gives us the comparison. Her actual behavior tells us how far beyond the comparison the story goes. https:// marvel.com/characters/dia mond-lil … https:// uncannyxmen.net/characters/dia mond-lil … https:// gia.edu/gia-news-resea rch/how-protect-diamond-chipping … https:// openstax.org/books/universi ty-physics-volume-1/pages/9-2-impulse-and-collisions … https:// marvel.fandom.com/wiki/Alpha_Fli ght_Vol_1_95 … https:// uncannyxmen.net/characters/dia mond-lil/checklist …

@
@loganpendragonm
@loganpendragonm

Behind the Science: Laura Dean Laura Dean has one of the most useful powers in this group, and it's not something you can measure by how hard she punches. She can open passages into other dimensions. Her sister Goblyn's history is tied to Liveworld, and their early appearances can make it look as though Laura is changing into Goblyn. She isn't. They are twins, and the passage between worlds is the important part. That difference matters. Shape-changing means rebuilding one body. What Laura does means connecting places that people normally cannot reach. The closest familiar science-fiction comparison is a wormhole: instead of crossing all the space between two locations, you take a shortcut connecting them. But that's a comparison, not a scientific explanation of her mutation. Liveworld is a fictional realm. We haven't discovered it, and we don't have evidence that a human brain can open a doorway into another universe. Even hypothetical wormhole models don't give you a practical instruction manual for building a safe door. Getting a mathematical possibility and getting a person through it alive are very different things. And the doorway itself raises questions the comic doesn't need to stop and answer every time. Are the air pressures the same on both sides? Is the destination safe to stand in? Does the traveler keep the same motion when they cross? What happens if the passage closes while something is halfway through? Those aren't extra powers I'm assigning her. They're the problems a real version of this ability would need to solve. What makes Laura so valuable is that transportation changes everybody else's options. An obstacle that stops the strongest member of the team may stop being relevant if she can provide another route. Escaping, retrieving someone, and reaching an otherwise inaccessible place can matter more than winning a fight. I also want to keep her separate from the idea that an unusual way of communicating makes someone less aware or less important. Her connection with Goblyn is central to understanding both of them. The phrase "another dimension" can also be misleading. In mathematics, a dimension is an independent direction or coordinate needed to describe something. In superhero stories, the same word often means a separate realm with its own environment and inhabitants. Those are not automatically the same claim. Saying that an equation uses extra dimensions does not demonstrate a livable world behind a doorway. And it certainly doesn't establish that Laura's mutation could connect to one. That's the difference between using scientific vocabulary and having a scientific mechanism. Still, imagining what a safe portal would require is useful. Consider the boundary itself. An ordinary doorway connects spaces that already sit next to each other. A dimensional passage needs an explanation for how its two sides become adjacent when the surrounding environments aren't. You could imagine a shortcut in geometry, a transfer of matter, or destruction and reconstruction. Those are three very different premises. The last of them would create questions about preserving the entire body's structure and whether the person arriving is continuous with the person leaving. I wouldn't attach that model to Laura without the comic saying so. A visible portal does not automatically imply scanning or copying. The safest comparison is a passage: someone crosses a boundary and arrives somewhere else. Even then, the boundary needs to handle more than the person walking through it. Air is a good example. If two connected spaces have different pressures, gas tends to flow from the higher-pressure side toward the lower-pressure side. A large mismatch could turn a convenient exit into a dangerous situation. That doesn't mean Laura's comic portals always have that problem. It means a real doorway between arbitrary environments would need to address it. You can put a number on the scale without pretending it measures her power. Suppose a sealed one-square-meter surface has a pressure difference of 10,000 pascals across it, around a tenth of normal atmospheric pressure. Pressure multiplied by area gives a force of 10,000 newtons. An open passage wouldn't behave like a solid surface resisting that force; the gas would move instead. The calculation illustrates why pressure matching matters, not the actual load on a magical portal. Temperature and chemistry matter too. A destination with breathable-looking air might still have the wrong composition. A safe crossing requires more than enough room for a person. In practical terms, knowing the destination could be as important as the ability to reach it. Movement creates another problem. Suppose a person is falling when the passage appears. Do they retain their speed after crossing? If so, where does that velocity point in the destination's coordinate system? If the portal stops them, where does their kinetic energy go? The story can establish any consistent fictional rule, but a science discussion shouldn't silently assume the most convenient answer. The same applies to moving worlds. A destination may rotate or orbit differently from the departure point. This is one of those invisible conveniences of portal fiction: the traveler usually arrives moving with the place they enter rather than suffering a catastrophic relative-speed mismatch. None of those questions establishes that Laura personally controls pressure, atmospheric chemistry, or velocity. They identify additional conditions a physical system would have to manage. Keeping that distinction prevents an interesting comparison from becoming a list of powers the character has never demonstrated. There is a planning side to her ability as well. A route can change the tactical problem without directly harming anybody. If you can bypass an obstacle, you may not need to destroy it. If you can remove someone from danger, you may not need to win the confrontation first. But useful does not mean unlimited. I would not convert "can open dimensional passages" into "can reach any place she wants instantly, with any number of passengers, under any conditions." Those are separate questions of range, destination control, capacity, and endurance. They should be tied to what the story actually shows. That restraint matters especially when discussing a character whose most important early relationship is personal. The door isn't just a convenient effect. It connects Laura to Goblyn and to the circumstances that shaped their lives. A power that looks like transportation can carry far more emotional weight than another energy blast. The physical explanation is firmly in the fictional side of this series. The practical importance is easier to understand. Being able to reach a place nobody else can reach can decide the entire situation before anybody throws a punch.

@
@Nfulanii
@Nfulanii

Most people look at @IPXExchange and see "IP as tokens." The part I find more interesting is where the money is supposed to go afterwards. IPX is building an operating system, launchpad and exchange for intellectual property on Base. Launch is scheduled for November 15, 2026. Planned categories include movie and TV projects, athlete video games and YouTube creator memberships, with fans able to buy, sell, trade and stake the project tokens. Here's the link that matters: every project token pairs with $IPX . So a creator's membership token and a film's token both route back to one shared asset instead of living in isolated markets. The proposed revenue loop, as IPX describes it: 1. Projects earn. Creators and IP owners bring content, products and communities. 2. Revenue flows back. IPX says it bridges traditional finance and blockchain, including non-custodial virtual checking accounts for deposits and withdrawals, and that project revenue carries contractual obligations to return through the system. 3. Buybacks follow. Listed sources: platform fees, IP revenue, sponsorship commissions, network fees, subscriptions and income from off-platform subsidiaries. A well-designed model isn't a proven one. What I'm watching: • Real revenue: buybacks need actual income behind them. Until projects earn meaningfully, the loop is untested. • Enforcement: some obligations rest on legal contracts, and smart contracts can't guarantee off-chain payments come back. • Concentrated risk: with every pair against $IPX , shared liquidity helps, but trouble in $IPX can spread across the ecosystem. • Transparency: I want records of how much was bought back, when, and funded by what. Announcing a mechanism isn't the same as executing it consistently. On FAN SZN: campaign details say top participants share part of the $IPX supply, unlocking linearly over 14 days. Check the official terms for exact allocations and conditions. The real question isn't whether fans will hold creator tokens. It's whether backing a creator can plug into a system where that creator's actual business creates measurable value for everyone else. The structure is laid out. Now the numbers have to prove it.

@
@_ummukulthum
@_ummukulthum

Wetin Really Connects $IPX , Creators and Fans? One thing wey catch my attention about @IPXExchange no be just the idea of turning intellectual property into tokens. Na how dem wan connect all the money wey those projects make back to one ecosystem. Make I break am down. IPX dey build an operating system, launchpad and exchange for intellectual property on Base, with launch scheduled for November 15, 2026. The plan cover things like movie and TV projects, athlete video games, and YouTube creator memberships. Fans go fit buy, sell, trade and stake the tokens connected to those projects. But the interesting part na the relationship between those project tokens and $IPX . Every project token dey pair with $IPX . Meaning say instead of every project standing alone with its own trading pair, $IPX go serve as the common asset across the platform. Why this matter? Imagine say one creator launch membership token and another project launch token for a movie. Both projects fit attract their own communities, but their tokens still connect back to $IPX through the trading pairs. Na one shared market structure. Now look at how IPX wan make the money move. The model get three important parts: Projects generate revenue. Creators and IP owners bring their content, products and communities into the ecosystem. Revenue returns through the platform. IPX say e dey connect traditional financial infrastructure with blockchain, including non-custodial virtual checking accounts for deposits and withdrawals. Dem also say project revenue get contractual obligations to flow back through the system. Platform revenue supports $IPX buybacks. Dem list different sources, including platform fees, IP revenue, sponsorship commissions, network fees, subscriptions and income from off-platform subsidiaries. So the bigger picture be say project activity fit generate revenue, some of that income fit return to the ecosystem, and part of the platform’s revenue fit support $IPX buybacks. For fans, staking and revenue-based rewards na part of the value proposition. But I no go confuse a well-designed model with a proven one. Four things I go dey watch closely: Actual revenue: The buyback mechanism need real money behind am. Until the projects start generating meaningful income, the loop never really prove itself. Revenue enforcement: Some of the obligations depend on legal contracts between IPX and project owners. Smart contracts alone no fit guarantee say every off-chain payment go return. Market exposure: Since all project tokens dey pair with $IPX , the condition of the native token fit affect the wider ecosystem. Shared liquidity fit create opportunities, but e fit also concentrate risk. Buyback transparency: I wan see clear records showing how much dem buy back, when dem do am and where the money come from. Announcing a buyback mechanism no be the same as demonstrating consistent execution. One more thing concern FAN SZN. Campaign details talk say top participants go share part of the $IPX supply, with allocations unlocking linearly over 14 days. So instead of receiving everything at once, the allocation dey release gradually. Participants suppose check the official campaign terms for the exact allocation and conditions. My own take? The interesting question no be simply whether fans go hold creator tokens. Na whether supporting a creator fit become part of a system where the creator’s actual business activity creates measurable value for the wider ecosystem. IPX get a structure wey dey try make that connection possible. But I go judge am by execution: projects wey generate revenue, money wey actually return, buybacks wey dem report, and rewards wey fans fit verify. The architecture don get explanation. Now na the real numbers go tell the story.

@
@arosplatforms
@arosplatforms

Every AI system we deploy has a cost per decision. We put a number on it before go-live, not after the first invoice. Here's how we run cost engineering on enterprise AI builds. We define the unit: a reconciled account, a triaged alert, a reviewed contract. Then we track the full cost of producing one, including model calls, retrieval, compute, human review time and rework. In design, we route simple steps to smaller models, cache repeated context, batch work that doesn't need to be real time and stop agents that loop. In production, finance gets a dashboard by use case and business unit, with budgets and alerts when cost per unit drifts. And we re-run the numbers when a new model lands. Sometimes a cheaper model meets the same quality bar on the client's golden set. Sometimes it doesn't. If you can't state the cost of one AI decision, you can't manage the return on a thousand. #LLMOps #FinOps #EnterpriseAI

@
@Gabrielcrypton
@Gabrielcrypton

当交易量放缓时,持有者奖励会发生什么? 这是一个值得询问的关于任何基于奖励的启动平台的疑问,包括 @RebateLaunch 。 Rebate Pad 的一个有趣特性是其自动持有者支付系统。 符合条件的持有者可以收到与其代币配对资产的分配,而无需手动领取奖励或质押。 但这个模型背后有一个重要的细节。 奖励取决于经济活动,而不仅仅是时间的流逝。 想象一个代币在其第一周产生显著的交易活动。 更多的符合条件的交易可以创造更多机会,让费用流入其持有者奖励系统。 现在想象同一个代币一个月后。 交易活动放缓,交易次数减少,可用于分配的金额可能会下降。 持有者并不一定做了什么不同的事情。 资助奖励的底层活动已经发生了变化。 这就是为什么我认为交易者应该超越启动平台奖励分配的频率来审视。 一个项目可能进行数百次支付,但这并不能自动告诉我们其奖励生成是否可持续。 在研究 Rebate Pad 上的代币时,我会关注三件事: • 交易活动是否保持一致。 • 实际分配给持有者的价值有多少。 • 随着初始启动兴奋消退,奖励是否仍然有意义。 并且记住,接收持有者奖励并不能消除代币价格下跌时亏损的可能性。 奖励机制可以分配价值,但它无法制造交易需求。 对我来说,这就是更大的教训。 有趣的问题并不仅仅是持有者多久被支付一次。 而是支持这些支付的活动是否能够持续下去。

@
@vitalsunapp
@vitalsunapp

A good vitamin D level doesn't mean you've gotten everything sunlight does for your immune system. ⠀ Vitamin D is just one thing UVB makes in your skin. Sunlight hits your immune system through a bunch of other pathways, and each part of the spectrum does its own job. ⠀ UV ⠀ UV calms down overactive immune responses by boosting regulatory T cells, the cells that stop your immune system from attacking the wrong things. In mice, UV suppressed a model of multiple sclerosis without raising vitamin D at all, and in people getting UVB therapy, some of the immune changes happened separately from their vitamin D levels. ⠀ Blue ⠀ Your skin holds around a million T cells per square centimeter, and they can sense light directly. In the lab, blue light made T cells move faster, and normal time outside is enough for sunlight to reach the ones near the surface. ⠀ Near infrared ⠀ This is the deep penetrating part of sunlight. It powers up your mitochondria and calms TLR4, a receptor that drives a lot of excess inflammation. ⠀ Morning sunlight ⠀ Your immune cells run on a daily rhythm too, from when they travel around your body to when they release signals. Morning sunlight in your eyes is what keeps that rhythm on time. ⠀ None of that shows up on a vitamin D blood test, and a capsule can't recreate it. Just keep it to regular, non-burning doses, since burning flips UV from helpful to damaging.

@
@flyinghigherai
@flyinghigherai

Before throwing the full 40-image dataset at the APIs, I ran a 3-image "smoke test" across our core archetypes: • 1 clean fungal case • 1 ambiguous sample with surface interference • 1 negative out-of-domain shot

@
@flyinghigherai
@flyinghigherai

Before throwing the full 40-image dataset at the APIs, I ran a 3-image "smoke test" across our core archetypes: • 1 clean fungal case • 1 ambiguous sample with surface interference • 1 negative out-of-domain shot

@
@flyinghigherai
@flyinghigherai

Before throwing the full 40-image dataset at the APIs, I ran a 3-image "smoke test" across our core archetypes: • 1 clean fungal case • 1 ambiguous sample with surface interference • 1 negative out-of-domain shot

@
@MikeTamir
@MikeTamir

TypeSafe AI 推出 Jev,这是一个仅限决策的模型,它跳过文本生成,直接返回类型化的概率决策。旨在用于快速路由、分类和代理预筛选。 https:// is.gd/85jcuX #MachineLearning #AI #LLM #DeepLearning #AgenticAI

@
@GitTrend0x
@GitTrend0x

Jev 应用分享又燃爆了 🔥 shitianfang 的 LLM + Jev 双人搭档、andududu 的 Jev 调用实况地图、tostechbr 的「话没说完就动手」语音控制、bytelabs-oss 的皇室战争机器人、AgriciDaniel 的技能分诊台……全网玩家把 Jev 玩成了下一代「快手搭档 + 调用监视器 + 读心语音助手 + 电竞选手 + 智能分诊员」: 1️⃣ shitianfang/jev-use( https:// github.com/shitianfang/je v-use …) 给 Claude Code、Codex 和 Pi 用的插件:不需要输出文字、只要快速判断的步骤交给 Jev,需要写内容时再交回大模型;作者实测拦截危险命令约 230 毫秒、不耗大模型 token,200 条消息 7 次调用完成上下文压缩;一行 npx 安装;72 星。 “写字交给大模型,拍板交给 Jev”! 2️⃣ andududu/jeview( https:// github.com/andududu/jeview) 一个非官方的本地 Jev 可视化网关:把 Jev 客户端指向它,它带着你的 key 转发请求,把每次调用存进本地 SQLite,并实时画在一张地图上;只需 Node 24,零依赖;作者声明与 TypeSafe 无关联;68 星。 “每一次 Jev 判断,都看得见”! 3️⃣ tostechbr/partway( https:// github.com/tostechbr/part way …) macOS 语音控制:你说的每个词都拿去问 Jev,约 200 毫秒返回带概率的判断,所以一句话还没说完,备忘录就已经打开了;能开应用、新建笔记、开网站、搜索、打字,还能一口气连说多个指令;32 星。 “你还在说,它已经做完了”! 4️⃣ bytelabs-oss/clash-jev( https:// github.com/bytelabs-oss/c lash-jev …) 用 Jev 在真实安卓手机上打《皇室战争》:OpenCV 把实时画面读成 JSON 局面,Jev 每次约 135 毫秒依次选策略、选卡、选落点,程序确认能下再点击;不用强化学习、不用模拟器,没有训练过任何出牌策略;37 星。 “没练过一局,照样上场对线”! 5️⃣ AgriciDaniel/gatekeeper( https:// github.com/AgriciDaniel/g atekeeper …) 在你的 AI 选工具之前,先决定这条请求该交给哪个 Agent 或技能:代码先过规则,剩下的语义问题交给 Jev 做带概率的选择,再按置信度给出放行、转交、建议、需确认、升级或拦截;每次不到一美分、约半秒;目前以 Claude Code 钩子安装;25 星。 “请求进门先分诊,别让 AI 瞎猜”!

@
@MazuzAsaf
@MazuzAsaf

Pi Coding Agent, plus Jev. PI-EXTENSIONS (NARUMIRUNA) Pi Coding Agent extension monorepo with experimental TypeSafe packages: pi-typesafe for typed Jev decisions, pi-typesafe-compact for compaction selection, and pi-typesafe-search for FTS5 plus Jev search. ~750 stars. When to use: you already work in Pi and want Jev handling compaction and search decisions. https:// jevlist.ai/projects/narum iruna-pi-extensions … More projects: https:// jevlist.ai

@
@kr4chinin
@kr4chinin

Currently building a project, where I’m deduplicating bug reports with an LLM. 11 hand-labeled cases, with traps. -> Cheap model: 9/11, ~150 ms -> Big model: 11/11, ~350 ms -> Cheap + fall back to big when unsure: 11/11 The fallback fired exactly twice, on the two cases the

@
@SuperagenticAI
@SuperagenticAI

🧠 SuperQode’s SystemOne harness now supports OpenAI @OpenAIDevs and Google @googledevs decision models alongside TypeSafe @typesafeai Jev. Available in SuperQode 2.31.17 and SuperQode for macOS 0.4.10. Watch the demo 📙 Full Blog: https:// super-agentic.ai/resources/supe r-posts/superqode-systemone-openai-google-decision-models …

@
@Shashikant86
@Shashikant86

🧠 SuperQode’s SystemOne harness now supports OpenAI @OpenAIDevs and Google @googledevs decision models alongside TypeSafe @typesafeai Jev. You can choose a decision model separately from your coding model. You can see what SystemOne recommends while your existing permission policy controls execution. Available in SuperQode 2.31.17 and SuperQode for macOS 0.4.10. Watch the demo 📙 Full Blog: https:// super-agentic.ai/resources/supe r-posts/superqode-systemone-openai-google-decision-models …

@
@fabrizio_degni
@fabrizio_degni

Jev didn't just launch a new AI model: it exposed a missing layer in AI Governance. They say one swallow doesn't make a summer but sometimes, one swallow signals an entirely new season. In September 2026, TypeSafe AI Jev introduced the System One paradigm: fast, structured, probability-driven judgments without traditional autoregressive text generation and the AI ecosystem responded almost immediately: Microsoft Decision-1. Gemini Decision 1. Perplexity Decider. OpenAI Decisions API. Cloudflare Clef. Liquid AI d1. Baseten Mammoth. Fastino Labs GLiDE. Amazon Web Services (AWS) Strands Decider. Within weeks, an entirely new category of AI intelligence was taking shape and the disruption isn't speed, it's what speed makes possible. When AI decisions become dramatically faster and cheaper, they can be embedded everywhere: in workflows, tool calls, authorization checks, agent interactions and autonomous swarms. Traditional governance was designed around systems, processes and periodic reviews. Agentic AI requires something more and dual approach: 🔷 TOP-DOWN: Translate organizational principles, risk appetite, accountability and delegated authority into operational boundaries. 🔷 BOTTOM-UP: Turn every consequential agent action into an opportunity for real-time verification, controlled execution, monitoring and auditable evidence. The transformation we're exploring with PALO Framework: 🧭 PALO Framework: Governance across the complete AI lifecycle. 🤖 PALO-AM: Authority, autonomy and accountability for AI agents and agentic swarms. 🛡️ PALO-AI: Enforceable controls, authorization gates and evidence of execution. 💰 PALO FinOps: Economic visibility into the cost, quality and sustainability of AI decisions. Together, they connect governance intent with operational reality: FRAME → CLASSIFY → ASSESS → CONTROL → MEASURE → PROVE & REVIEW Decision Models can provide a powerful, low-latency judgment layer within this pipeline but they must never become the ultimate authority. The frontier of AI Engineering is building AI ecosystems in which every action has a defined authority, every risk has an enforceable boundary, and every consequential outcome can be verified. 📊 The infographic below maps the emerging Decision Model landscape as of November 2026, its strengths and limitations and its implications for AI governance. 🌐 Check it out here and on the 1st comment: https:// paloframework.org #PALOFramework #AIGovernance #DecisionModels #AgenticAI #AISwarms #ResponsibleAI #AIEngineering #SystemOne

@
@SJRonanMD
@SJRonanMD

Zhejiang and Baidu: a decision model that acts badly can still teach an agent well Zhejiang University and Baidu found that a decision model that's bad at running an agent can still be good at training one, as long as you ask it the right question. My take: JEV, a closed decision model from TypeSafe AI that returns a choice with probabilities instead of text, completes 42.9 percent of ALFWorld tasks when it picks the actions itself. Used as a frozen judge that scores every action of a Qwen2.5-1.5B policy after the episode ends, it lifts GRPO training from 72.8 to 82.9 percent on ALFWorld and from 56.8 to 71.4 on WebShop. The wording matters. Asking whether each action helped worked. Asking after each action whether the task will succeed gave 69.3 percent, worse than plain outcome rewards. Their explanation is that actors get pulled toward an easier subgoal, while a judge that sees the whole trajectory can tell which steps mattered. JEV is closed and its base model isn't disclosed, it's two simulated environments, and the student is 1.5B. An open-weights decider from Perplexity is in the actor and judge comparison, but only JEV was used as a teacher. Hindsight judging looks like a cheap way to train small agents you'd run

@
@MiaSardinella
@MiaSardinella

两位内阁秘书解释为什么所有飞机都被停飞。他们在说同样的事情吗?👀

@
@Dr_ztc
@Dr_ztc

🚀 TypeSafe 估值不到一个月涨37倍:约20亿→740亿瑞典克朗($75亿) 它做什么:旧金山AI公司,2024年成立,联创为前OpenAI研究员Diogo Almeida;做给软件直接调用的AI模型,首款Jev主打自动化客服等高频决策。 钱:$8.7亿A轮,a16z领投、红杉参投;Klarna CEO旗下Flat Capital跟投约1000万克朗。

@
@INTJYAN
@INTJYAN

🚀 TypeSafe 估值不到一个月涨37倍:约20亿→740亿瑞典克朗($75亿) 它做什么:旧金山AI公司,2024年成立,联创为前OpenAI研究员Diogo Almeida;做给软件直接调用的AI模型,首款Jev主打自动化客服等高频决策。 钱:$8.7亿A轮,a16z领投、红杉参投;Klarna CEO旗下Flat Capital跟投约1000万克朗。 观察:资金在抢"非聊天式"AI模型。 #资金动向 #人工智能 #AI #VC

@
@jabowery
@jabowery

Me❯ Some interventions may be a kind of capital investment to push the system into a different basin of attraction with higher utility, by using high leverage (low cost) inputs. These are the kinds of interventions that institutions tend to resist since their institutional

@
@monikayadav_3
@monikayadav_3

The biggest constraint on your business growth isn't your product or market - it's your content operation. You're trying to scale a one-person show in a world that rewards systems. We've analyzed over 200 founder brands. The pattern is clear: founders who DIY their content plateau fast. They hit 200-300 followers and stay there. Engagement stays below 50 likes. Inbound leads are non-existent. We see founders with 10K followers getting only 20 likes because the audience doesn't care about random posts. It's organized noise without strategy. Not because the content is bad, but because there's no engine behind it. Your time is finite. Spending 10-15 hours a week on content means less time on sales, strategy, or leadership. The math breaks down. You're trading your highest-value hours for mediocre distribution. We designed a different model. For a consultant in professional services, we built a full system: ICP research, competitor hook mapping, content batching, and automated posting. He approved posts in 20 minutes a week. We handled everything else. Results in 90 days: from 80 followers to 12,000. From 5 likes per post to 2,500. Profile visits jumped from 10 to 450 per day. He closed three inbound deals worth five figures each. This is how you scale without burning out. You invest in a system that compounds, not in hours spent writing. The agency becomes your growth engine - silent, systematic, and relentless.

@
@Andy_Luigino
@Andy_Luigino

Roadmaps are easy to read as reality, and the first time I opened the AINFT whitepaper, I mixed up what was already built with what was still a plan. That is an easy mistake, so I went back and read Whitepaper V2.0 from September 2025 slowly, with one question in mind: what is on paper, and what can I actually see today? Here is the paper version, four business lines: ➜ AINFT Agent Framework, a multi-agent system with Core API message passing, event-driven scheduling, plus AgentChat, Extensions and Magentic APIs ➜ AINFT Nova, a no-code agent launch platform where each agent can issue its own token, "one agent, one economy" ➜ AINFT AgenTX, a FinAgent framework for medium-to-low frequency multi-chain DeFi with ML forecasting, reinforcement learning and DAO-governed parameters ➜ AINFT Grid, decentralized model training on distributed compute with smart-contract task scheduling and incentives Underneath sit the roadmap layers: ➜ An AI Agent Layer with DID registration ➜ An SDK Layer and a User Interaction Layer ➜ An Incentive Layer and the Agent Commerce Protocol That is an ambitious stack, and I like that it is specific rather than vague. Now the honest part. An independent Protos review in February 2026 said Nova, Agent Framework, AgenTX and Grid had not launched at that time. What I can see on the AINFT site today is narrower: the AI Trading Agent Benchmark, and even that is labelled coming soon, private preview. So my gap analysis looks like this: ➜ Designed: 𝗳𝗼𝘂𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁 𝗹𝗶𝗻𝗲𝘀 with real technical detail ➜ Visible: one evaluation product, on BTC and ETH paper markets only ➜ Missing: public launch dates for Nova, Agent Framework, AgenTX and Grid I do want to give credit where it is due. The benchmark is not random; it maps neatly onto AgenTX, because you cannot responsibly run a trading agent before you can measure one. The TRON H1 2026 Strategy Report also places AINFT in the trading-strategy validation layer, which tells me the focus has narrowed on purpose. 𝗡𝗮𝗿𝗿𝗼𝘄𝗶𝗻𝗴 𝗶𝘀 𝗻𝗼𝘁 𝗳𝗮𝗶𝗹𝗶𝗻𝗴. Plenty of whitepapers try to ship everything at once and ship nothing well. My open question is sequencing: does Nova come before AgenTX, and will Grid wait until there is real demand for decentralized training? If the team publishes dates and milestones for each line, this whitepaper becomes a checklist we can all track together. Read it yourself and tell me which line you think should ship next: https:// ainft.com/whitepaper/AIN FT%20White%20Paper.pdf … 𝗣𝗮𝗽𝗲𝗿 𝗶𝘀 𝘁𝗵𝗲 𝗽𝗿𝗼𝗺𝗶𝘀𝗲; 𝘀𝗵𝗶𝗽𝗽𝗶𝗻𝗴 𝗶𝘀 𝘁𝗵𝗲 𝗽𝗿𝗼𝗼𝗳. @justinsuntron @AINFTcom #TronEcoStar

@
@JarryJeremy
@JarryJeremy

X 已将复制链接分享的排名权重从 20 降至 0。 这此前是已公布模型中权重最高的正面信号之一。 以下是 10 月 27 日至 11 月 15 日期间的其他变化: • 点赞:+7.3% • 一般分享:+5.9% • 连续阅读时间:+25% • 帖子点击:-8.7% • 点击相关阅读时间:-3.8% • 视频打开:降至零 • 复制链接分享:降至零 • “不感兴趣”:负面权重略微增强 几点值得了解的事项。 1. 阅读时间正变得越来越重要。 X 已在 10 月更新了其检索系统,以更好地识别用户完整阅读的帖子。 现在,连续停留时间系数又增加了 25%。 对于创作者来说,这为有实质内容的帖子提供了有力理由。一个强有力的开篇能吸引注意力,但其余部分需要值得一读。 2. X 正在改变其使用你浏览历史的方式。 最新的代码使排名服务能够接收多达 50 个先前提供的帖子的信息。 这为系统提供了更多关于某人已看过内容的上下文,并可能帮助其避免重复推荐。 对于创作者来说,反复发布相同的回收观点可能会变得不太有效。 3. 分享仍然重要,但并非所有分享都平等。 复制链接信号降至零尤为引人注目。 与此同时,一般分享的权重有所增加,而回复、引用和 DM 分享仍保持重要的正面信号。 一个重要的区别:这些权重应用于预测行为,而非原始互动计数。一个回复并不自动等同于五个点赞。 4. 算法越来越专注于预测有意义的互动。 我对这些变化的解读是,X 更加强调人们是否真正想花时间阅读帖子、讨论它并分享它。 不再有理由只痴迷于点击量。 一个注意事项:这些是公共代码默认值。X 可以运行不同设置的实验,因此我们不能假设每个用户看到的排名行为完全相同。 如果你想深入研究,完整源代码是公开的: https:// github.com/xai-org/x-algo rithm … 好奇其他人是否注意到过去几周他们的覆盖范围发生了变化。

@
@DerickZill
@DerickZill

OK before anyone gets too serious on this one... This is a real account of my personal trip and I was WAAAAY to tired to do this safely and would not advise anyone else repeating this. I myself never want to do it again and agree that it is stupid. I probably almost fell asleep 7 times and this trip would not be possible without FSD. ******************************* (DO NOT DO THIS) ******************************* It was dumb of me to do it and I will not be doing it ever again if I can help it. I recently picked up my New 2026 Model Y from Tesla and honestly, it’s a pretty awesome car. Here's the pro’s con’s and other info. I Left Dryden at Midnight November 10th My father drove me to Winnipeg a ~4h drive for a 2.5 h flight to Toronto. I arrived in Toronto, then took a 15 minute cab ride down to Etobicoke for the pickup. (Yes I asked multiple times and they won't deliver closer... FML...🤦) My appointment was scheduled for the 11th month on the 11thday at the 11th hour. 11,11,11 :D I arrived around 9 and the Tesla Center wasn’t open until 10. The cab driver had dropped me off at the front door inside the fenced area. This was amusing because everything was closed and the only exit was where he drover in. Basically I was early so I awkwardly walked around and looked at some cars before realizing I probably wasn’t supposed to be in there yet. I left to go get a drink and went back at about 10. They let me go in and start my paperwork right way. The staff at the Tesla Center were pretty good. He showed me the car, ran through the features, tested out FSD and talked about a few other issues or concerns I had. I asked about accessories and was a little disappointed in the lack of stuff available on site. I did pick up a charging cord for my home, but there wasn’t much else there for accessories at all. I was honestly hoping for a show room or at least an air mattress or something. Anyways... There was an Optimus at least 😅 By about ~11:30 I was off for my very very very long drive home... I let FSD take over almost the entire drive home and it actually did surprisingly well. It navigated amazingly, parked itself most of the time, and had very little issues in total. "This car is amazing at city driving" It's basically like taking a cab, country driving needs some work still. 🤷 Here are the few issues it did have. Please note that these issues occurred over the span of 1700+ km and were not present the whole time. I was waaaaay to tired to be driving and was definitely way more of a safety risk than FSD could ever be. These issues were only present for short periods during the entire 1700+ km drive. 1) Speed limit signs. The FSD system seemed to have a problem with not seeing or completely disregarding speed limit signs in certain areas. This happened mostly outside of the area before Blind River, and then again much closer to home. For a vision based system, I am unsure how it could miss 3x speed change signs multiple times? There is usually a Speed sign with an arrow signifying the speed limit will change up ahead, another sign without an arrow signifying this is where the speed changes, followed by a third sign just in case you missed the first 2... Yet the Tesla would stay going 90 when it dropped to 60 or 70kpm or insist it was still 60 or 70kph when it went back up to 90. This should probably be fixed... There is also the yellow signs with sharp angles and advisable speeds for sharp corners and stuff, yea.. FSD should probably pay attention to those too... Nothing wakes you up like a 90 degree right arrow that says like 30kph and the cars like fuck that, lets do 90 🤣 (I Suggest adding a cruise toggle or control feature where you can manually set the desired speed and FSD will try and maintain it) Maybe just have it on the screen or use one of the toggle wheels on the steering wheel so you can adjust it like many adaptable cruise control systems. This would allow people to adjust it to the proper speed without having to jerk it out of FSD and the car could learn from the consistency across the network. 2) Freezing Rain and Icy road conditions. Yes I know people are probably going to say, “you shouldn’t use FSD in those conditions” “this is an unfair test” so forth whine whine. The bottom line is this. FSD will eventually drive everyone everywhere in all road conditions, there's no point in babying it if you want it to learn. FSD actually did really good, there were only two major issues and one of them may have been related to ice buildup in the frontal camera combined with someone else's trailer or something. Firstly, FSD NEEDS to be able to assess icy road conditions and adjust brakes accordingly. Sudden or rough braking on ice is just asking for problems. The car still handled fine, but when approaching another vehicle going 20kph slower due to icy road conditions, maybe don’t use the brakes last second so that you skid up behind them? it’s not city driving so the car should adjust accordingly. Slow deceleration, slow into the corners and faster out, no sudden speed changes, you know, icy road driving... The second was a rare instance of what can only be described as a "death wobble". Basically the car came up behind another vehicle with a tow load. The car then decided that going slightly left, then right, then slightly more left then right, then slightly more left then right, in a slowly increasing fashion on an icy road was a good idea? Yea... it is NOT...😐😅It is a good way to test the vehicles stability in a small recovery skid though... FYI it’s incredibly stable due to it’s low center of gravity and could probably drift a 90 degree corner at 100km an hour if you wanted to try it. (not recommended though 😅) I actually needed to disengage FSD on this one because it would have gotten dangerous fast... This could have been due to the frontal camera being iced up a bit, or the car trying to recover from a slide/skid coming in behind the other vehicle. Regardless, it needs to be looked at for sure. 3) parking panic and then just not caring. FSD kinda panicked when trying to park for charging when all the lines were covered up, it’s ok but a little amusing that a light dusting of snow cripples the parking feature. Slightly later it didn’t seem to care about the lines or lining up at all, I actually had to adjust the parking manually just so the charge cord would reach, I found this a little odd. Maybe it was related to loss of signal or something, I know cell reception out this way is often non existent for massive areas. I find this behavior odd as it usually parks itself perfectly, I was actually really impressed by the smoothness of it when I picked it up. Suggestion. Maybe have it line up with the charging stations themselves? It can stull use all the obstacle avoidance features of the driving, but have it line up with the charge station when the lines aren't available. This would allow people to still line up for charging even if the lines are a problem. Overall FSD gets 4.9 Stars out of 5. The car passes, navigates, drives, avoids obstacles (dodged a tire on the road) parks, and does everything its supposed to do. I am still testing it and considering looking for some studded winter tire options. Overall it’s really really damn good. I arrived home just before 10:30 am November 12th, this means I was physically up and awake for ~35 hours. 23 of these hours were in the drivers seat with only short charging naps, usually ~15 minutes or so, this thing charges FAST! Again I “DO NOT” recommend doing this... (I probably almost fell asleep 7 times 😑) DON’T DO THIS!!! I DO recommend having that air mattress option in stock and really pushing sales for anyone who plans on driving more than 5 hours on pickup... If anyone has any questions, feel free to ask, I’m more than happy to go find out some new stuff too if I don’t yet know.

@
@Crypt_cap
@Crypt_cap

𝗨𝗦𝗗𝗗 𝗜𝗦 𝗠𝗢𝗥𝗘 𝗧𝗛𝗔𝗡 𝗔 𝗗𝗜𝗚𝗜𝗧𝗔𝗟 𝗗𝗢𝗟𝗟𝗔𝗥. 𝗜𝗧'𝗦 𝗔 𝗗𝗘𝗖𝗘𝗡𝗧𝗥𝗔𝗟𝗜𝗭𝗘𝗗 𝗙𝗜𝗡𝗔𝗡𝗖𝗜𝗔𝗟 𝗜𝗡𝗙𝗥𝗔𝗦𝗧𝗥𝗨𝗖𝗧𝗨𝗥𝗘. When evaluating a stablecoin, its price peg is only one part of the conversation. The governance model, transparency, security architecture and practical utility are equally important. This is where USDD ( @USDDio ) deserves a closer look. Built around decentralization and user empowerment, USDD brings together several features designed to support a more accessible and transparent stablecoin ecosystem. Here are six important things to understand: 🔹 Decentralized Governance: USDD incorporates community-driven governance through proposals and on-chain voting, giving governance participants a role in shaping the ecosystem. 🔹 Transparency & Auditability: Blockchain records make transactions and relevant collateral information publicly verifiable, allowing users to examine important aspects of the system. 🔹 Over-Collateralization: USDD uses an over-collateralized reserve framework designed to provide additional backing and help manage market fluctuations. This does not eliminate market or protocol risks, but it is an important part of its stability architecture. 🔹 DeFi Integration: USDD can participate in decentralized financial activities such as lending, borrowing, savings and trading across supported applications. 🔹 Freeze-Free Transactions: USDD held in a self-custodial on-chain wallet cannot be directly frozen at the token level by a centralized authority. Users still need to consider smart contract, wallet and network risks. 🔹 Multichain Accessibility: USDD currently supports TRON, Ethereum and BNB Chain through its application, expanding its reach across different blockchain environments. But here's the bigger picture. A useful stablecoin should offer more than a dollar-denominated value. It should provide accessible infrastructure, transparent mechanisms and practical opportunities within the digital economy. USDD brings these elements together in its approach to decentralized finance. Understanding these fundamentals is an important step toward understanding where decentralized stablecoins fit into the future of finance. @justinsuntron @USDDecentralize #TRONEcoStar

@
@maxt3chno
@maxt3chno

批评者说“没人用 Three.js 开发” 好吧……解释一下这个: 一个编码模型怎么能和 Blender 中烘焙的 3D 网格相媲美?? 顶部是 Three.js 底部是 Blender 水、天空中所有的模拟机制 = 同样通过 Opus 5.5 + Fable 5.2 构建 有疑问吗?

@
@OluJoe100
@OluJoe100

⚖️ NIGERIA POLITICS DEBATE: Bola Tinubu (Lagos) vs. Peter Obi (Anambra) — Who Left the Stronger Record as Governor? As campaigns heat up ahead of the January 2027 presidential election, one of the biggest debates dividing Nigerians online and at campaign rallies is the comparison between Bola Tinubu’s two terms as Governor of Lagos State (1999–2007) and Peter Obi’s two terms as Governor of Anambra State (2006–2014). According to political analysts, party spokespersons, and public governance debates across Nigeria, both camps point to very different leadership styles and achievements: * The Case for Bola Tinubu in Lagos (1999–2007): Supporters and APC spokespersons point to Tinubu’s financial engineering and long-term master plan for Lagos. When former President Olusegun Obasanjo withheld Lagos local government funds, Tinubu’s administration overhauled the state's tax system—boosting Internally Generated Revenue (IGR, meaning money the state raises locally without waiting for Abuja) from a few hundred million naira monthly to billions. His backers also credit him with laying the institutional blueprint for Lagos’s modern transportation (BRT), waste management, and power projects, while mentoring a team of successors who carried those plans forward. • What critics say: Opponents argue that Lagos already enjoyed a massive head start as Nigeria's former federal capital and commercial hub, and question the high cost of governance and debt profile left behind. * The Case for Peter Obi in Anambra (2006–2014): Supporters and public development commentators point to Obi’s strict cost-cutting, prudence, and focus on human capital despite Anambra’s much smaller revenue base. Backers highlight his partnership with the church to hand back and revitalize mission schools—which pushed Anambra to the top of national WAEC and NECO rankings—alongside expanding rural road networks and leaving billions of naira in cash and investments in state reserves without borrowing from banks. • What critics say: Presidency spokespersons and critics argue that saving money in banks rather than spending aggressively on mega-infrastructure held back faster industrial growth, and question his succession planning and local government elections during his tenure. Let’s keep the comment section respectful and factual: Which governance model do you believe Nigeria needs most right now—Tinubu’s revenue-and-infrastructure expansion or Obi’s cost-cutting and human-capital focus? Drop your thoughts below! 👇🏾 #NigeriaPolitics #NigeriaElections2027 #BolaTinubu #PeterObi #LagosState #AnambraState #NaijaDebate #NigerianYouth #Election2027 #NaijaNews

@
@Crypt_cap
@Crypt_cap

𝗧𝗥𝗨𝗦𝗧 𝗜𝗡 𝗗𝗘𝗙𝗜 𝗦𝗧𝗔𝗥𝗧𝗦 𝗪𝗛𝗘𝗡 𝗨𝗦𝗘𝗥𝗦 𝗖𝗔𝗡 𝗩𝗘𝗥𝗜𝗙𝗬 𝗪𝗛𝗔𝗧 𝗧𝗛𝗘𝗬'𝗥𝗘 𝗧𝗥𝗨𝗦𝗧𝗜𝗡𝗚. One question every stablecoin user should ask is simple: Can I independently examine the system responsible for protecting the value of the asset I hold? This question matters because a stablecoin's credibility should not depend entirely on what its developers say about it. Its underlying mechanisms should also be open to inspection. This is an important part of the infrastructure behind USDD ( @USDDio ). Let's look at three areas where transparency becomes practical rather than just a marketing promise. 🔹 1. Visible Collateral Infrastructure USDD uses a collateral-backed model. Its official documentation provides contract addresses that users can examine through supported blockchain explorers. This gives participants a way to inspect relevant on-chain balances and understand where protocol collateral is held. 🔹 2. Transparent Protocol Operations Through its Vault system, users can examine collateral positions, outstanding debt and collateralization ratios. These details matter because maintaining a healthy collateral position is essential to managing liquidation risks and supporting the protocol's financial structure. 🔹 3. Accessible Peg Stability Mechanism USDD's Peg Stability Module (PSM) provides a conversion mechanism between USDD and supported stablecoins at a fixed 1:1 rate. Rather than relying entirely on conventional decentralized exchange price curves, the PSM provides another route for stablecoin conversion and supports the protocol's peg-management approach. 𝗧𝗛𝗘 𝗕𝗜𝗚𝗚𝗘𝗥 𝗣𝗜𝗖𝗧𝗨𝗥𝗘 Transparency is not simply about publishing information. It is about giving users the tools to investigate, understand and evaluate the infrastructure behind their assets. Of course, on-chain visibility does not eliminate smart contract vulnerabilities, market volatility or collateral risks. Users still need to conduct independent research. But when important protocol information can be examined rather than accepted blindly, users have a stronger foundation for making informed decisions. That's an important principle for the future of decentralized finance: don't just ask users to trust the system. Give them the ability to verify it. @justinsuntron @USDDecentralize #TRONEcoStar

@
@theintgap
@theintgap

As more observers apply the Three-Track Model to contemporary relationship dynamics, a natural question arises: once we recognize that removing post-agricultural Institutional Scaffolding has exposed the raw Biological Floor, what is the prescriptive solution for the future? First, a strict methodological boundary must be stated. The purpose of this work is not to serve as a normative guide to the future or to prescribe a political or cultural reconstruction. In empirical science, diagnostic cartography must remain separate from ideological prescription. My sole objective is to map the present landscape with maximum accuracy so that observers can read the environment as it actually functions. Second, the confusion surrounding modern mating chaos disappears the moment we recognize that the Three-Track architecture governs all domains of human cooperation, not just reproduction. 1. The Universal Architecture of Social Technology In Dual-Inheritance Theory (gene-culture coevolution), researchers such as Peter Richerson and Robert Boyd (2005) and Ara Norenzayan et al. (2016) have demonstrated that large-scale post-agricultural civilizations could not function solely on raw Pleistocene instincts. Across every domain of survival, human behavior is sorted across three distinct bins: The Biological Floor: Conserved, self-interested evolutionary algorithms selected to maximize individual inclusive fitness and survival under ancestral conditions. Institutional Scaffolding: Co-evolved legal, religious, economic, and reputational technologies that impose external costs on floor defection, making large-scale cooperation and long-term contracts stable. The Cultural Canopy: The idealized moral or romantic narrative that a society layers over its institutions to describe human cooperation in flattering terms. 2. The Criminology Analogy: Why Only Mating Surprises Us Evolutionary psychologists Martin Daly and Margo Wilson (1988) and David Buss and Todd Shackelford (1997) have extensively documented that the human Biological Floor contains conditional, context-dependent modules for resource expropriation (theft, raiding) and interpersonal aggression. When the expected payoff of seizing another person's resources exceeds the expected cost of punishment, the floor strategy activates. In the domains of property and violent crime, society does not confuse the Cultural Canopy with the Institutional Scaffolding. No rational policymaker assumes that human civic politeness is self-sustaining, and no one proposes abolishing locks, police enforcement, property titles, and criminal penalties on the assumption that "mutual respect" will prevent burglary. If a city dismantled its legal scaffolding on Monday and experienced widespread looting on Tuesday, no observer would diagnose the population with a mysterious new psychological disorder. They would recognize that removing external constraints exposes baseline opportunistic behavior. 3. The Mating Exception and the Collapse of the Rails In the domain of human reproduction and pair-bonding, however, contemporary culture committed that exact error. Post-agricultural monogamous marriage was a high-constraint social technology built to align male provisioning, female hypergamy, and paternity certainty by pricing out opportunistic floor behavior (Henrich, Boyd, & Richerson, 2012; Fortunato & Archetti, 2010). Over time, our culture mistook the decorative Romantic Canopy (the story that soulmate emotion holds families together) for the load-bearing structure. Operating under that illusion, society systematically removed the legal, economic, and reputational scaffolding around pair-bonding while leaving the biological hardware unchanged. Today, we watch unfiltered Strategic Pluralism (Gangestad & Simpson, 2000), partner-specific erotic habituation (Klusmann, 2002), the Coolidge effect (Symons, 1979), modular partition defense (Kurzban & Aktipis, 2007), and rational male market withdrawal (Baumeister & Vohs, 2004) execute in the open. There is no mystery in the contemporary relationship landscape: while historical scaffolding was often harsh and punitive, demolishing an archaic constraint system without installing a viable replacement does not eliminate human conflict; it simply strips away the institutional regulation of an ancient biological drive and defaults the market back to the raw evolutionary floor.

@
@Mukshud_ahamed
@Mukshud_ahamed

A machine learning model that can steal information from your entire computer system by hijacking the physics of your computer chips? That’s what I was working on before my devices were temporarily magnetized and my uh GPUs got destroyed. Mostly motivated by cybersecurity. If one can think it then one can model it. It was part of an open source antivirus package I wanted to release as part of my PhD thesis project(s). @NSACyber

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

令人惊讶的是,很少有人意识到状态是共享的,而提出大量问题是非常高效的。

@
@EverymansAI
@EverymansAI

👉 S1 Watch #11: Google just joined the S1 race. Your release gate shouldn't change.💥 🙋 What's new Google DeepMind released Gemini Decision 1 and a Gemini Decisions API. Like Jev, it writes no prose. It returns typed answers with confidence scores your code can act on, and it reads text, images, audio, video and PDFs together, up to 1,000 questions about the same material in one pass (VentureBeat). Google joins the list this S1 Watch series tracks: ⇒ Jev, ⇒ Microsoft-Decision-1, ⇒ OpenAI's Decisions API, ⇒ Perplexity, ⇒ Liquid AI, ⇒ Drex, and now Gemini. 🤔 What's exciting 1️⃣ Multimodal checklists. A claim with a form, photos and a call recording can be screened in one request. 2️⃣ Independent referees are arriving. Artificial Analysis's Decision Index v2 rewards calibration, not just accuracy. 3️⃣ The interface is standardizing. As Pete Johnson put it in CIO, the interface standardized before the market did. That's what makes S1s swappable. Where I'm skeptical 🔸Self-run benchmarks. In Google's own launch tests, Gemini led five of seven. On Artificial Analysis's independent index, it launched ninth of 18. Same model, different referee. 🔸Right answers aren't a right case. That index's document test only counts a document when every answer on it is correct. The whole case is what you release. 🔸A probability is not a decision. Even TypeSafe says it plainly: "Jev makes judgments, code makes decisions." 🔸 Inputs can be steered. VentureBeat reports a single planted field dropped Jev's probability of blocking an action from 0.76 to 0.48. A model that can be talked down should never be what releases an answer. 🛠️ How we handle that in ModelFusion ▫️ Jev runs in production in our document lane, where it can discard a bad contributor draft. It filters. It doesn't release. ▫️ Only the deterministic audit gate can release an answer. ▫️ Every run leaves unsigned inspection receipts: what was checked, and why it passed or failed. ▫️ Gemini joins the rest on our watch list, not integrated. Google's entry confirms there won't be a single winner. What's your S1 bet: pick one lab and hope, or keep the gate yours and swap the scorer? DM me "demo" 🌐 http:// modelfusion.ai ----- AI Reliability Engineering: ✍️ Join the ModelFusion™ builder Linkedin group: https:// lnkd.in/d_68hMT4

@
@Gravipy
@Gravipy

嘿 @grok ,宇宙中最可怕、却几乎没人知道的东西是什么?

@
@sanderzag
@sanderzag

我曾在最大的大公司、小到不能再小的公司(我和兄弟创办的公司)以及介于两者之间的其他公司工作过。我无法预测未来,但我会问一个问题。 为什么那些议价能力最强的客户——GS、MS、KKR、Carlyle 等——总是坚持选择最昂贵的律师事务所?他们有很多选择,其中许多都能为他们节省大量资金,而且他们谈判时也毫不怯场。那么,为什么呢? 一旦你弄清楚这一点,你就会明白为什么 AI 不太可能颠覆基本的律师事务所模式,就像 LegalZoom 未能做到这一点一样。它 WILL 做的是推动系统内的效率提升,这将表现为大公司获得更多交易、更长的交易文件,以及顶级公司获得更多利润。

@
@joelbqz
@joelbqz

在 https:// sauv.io 画布中使用浏览器节点的情绪板

@
@Vijaya_1303
@Vijaya_1303

I replaced an LLM with Jev for product verification and reduced matching time from 45–60 seconds to under 20 seconds. While building BrandCompass AI, I realized I was using a text-generating LLM for a task that didn't require text generation. Product verification is a classification problem: YES / NO / UNSURE Instead of generating explanations I never used, I switched to Jev by TypeSafe AI, a decision model that selects from predefined options and returns a confidence score. What changed: Qdrant retrieval remained unchanged. Only the candidate verification step was replaced. Jev returns decisions instead of free-form text. Billing is based on input tokens; output is free. Result: Matching time dropped from 45–60 seconds to under 20 seconds on my test set. A wrong match is still possible, so sellers confirm every group. Next: moving matching to a background task with a progress bar. Would you use a decision model for pipeline classification or an LLM with structured output? #BuildInPublic #AIEngineering

@
@snaga
@snaga

Jev自体も面白いと思うけど、Jevを作り出したTypeSafeの人たちの頭の中にある世界観、みたいなものに興味がある。 何を、どのように見ているのか。それが「ビジョン」と呼ばれるものだと思う。 copycatに興味がないのはそういう理由。(実用上の価値は認めるとしても)

@
@rafaellinscto
@rafaellinscto

Um LLM escreve token por token e às vezes abre com "Claro, aqui está o JSON:", que quebra o parse. O Jev, da TypeSafe AI, não escreve: devolve decisões tipadas com probabilidade, direto para um if (números da empresa). https:// linspartners.com/guias/ferramen tas-de-ia/decisao-tipada/ …

@
@Ashtonexi
@Ashtonexi

判断题交给 𝗝𝗘𝗩-𝟭.𝟭𝟯.𝟬:它不写文章,只给你选项、分数和把握度。 http:// B.AI 文档里有一个很特别的模型页,是 TypeSafe AI 的 JEV-1.13.0。 按文档说明,它是 System One 模型,专门做软件流程里的快速结构化判断。 你定义问题和可选答案,它返回选择、分数或概率,由程序决定下一步。 文档列出的输出类型: ➜ 𝗖𝗵𝗼𝗶𝗰𝗲:从你给的选项里分类 ➜ 𝗦𝗰𝗼𝗿𝗲:按评分标准打分 ➜ 𝗡𝗼𝘂𝗹:估算是或否的概率 Choice 和 Score 还会返回概率分布和一个推导出的置信值,方便你处理模棱两可的情况。 多个独立问题可以共享同一份状态并行评估,减少串行调用。 版本方面,文档建议用 jev-1.13.0 固定稳定版本。 jev-latest 目前指向同一个模型,但以后可能变化。 适合的场景包括请求路由、工单分流、检索结果过滤和内容审核。 我的理解是,让大模型做创作,让这类模型做裁判。 完整页面在这里: https:// docs.b.ai/zh-Hans/llmser vice/models/jev-1.13.0/ … @justinsuntron @BAI_AGI #TronEcoStar

T
TypeSafe AI@typesafeai
@typesafeai

不问问题,就得不到答案!先上车,之后再搞清楚。 来自 @allietheicon 的最新最佳实践分享

@
@bitcoinduke
@bitcoinduke

it seems like every big lab is launching its own decision model, and people are actually using them decision models went from zero Vercel ai gateway requests in Aug to 26% in October (just 0.53% of total spend tho). almost all that volume went to typesafe ai’s Jev, but it helps explain why everyone seems to be launching their own now we’ve got Tev1 from Together AI, Clef from Cloudflare, and Gemini Decision 1 from Google. and it makes total sense, because these models are actually useful: when your app needs to classify something, score it or pick from a set of options, a specialized model can handle that without generating a whole response. it’s just impressive how fast this is happening

T
TypeSafe AI@typesafeai
@typesafeai

模糊的田野照片?Jev 迅速做出判断来查明真相。 @elise_ai 记录 Jev 在 94ms 内完成判断:一张照片的速度比 Sol 快约 30 倍,在相同成本下判断数量多约 50 倍

@
@pizuricv
@pizuricv

Proud father! congrats @marpizurica . RLCD in action https:// lnkd.in/p/eXx3PR6d MANTA, a family of multimodal decision models that answer structured medical questions with calibrated probabilities, available in two sizes (27B and 4B parameters).

@
@Mikeofweb
@Mikeofweb

📢 http:// B.AI POPULAR MODEL DISCOUNT UPDATE: GLM-5.3-FLASH NOW 40% OFF! http:// B.AI’s model discount lineup has officially been upgraded! The latest update brings even better deals on popular AI models, giving developers more opportunities to explore advanced AI capabilities at a lower cost. The biggest highlight is GLM-5.3-Flash, which has received an upgraded discount from 30% OFF to 40% OFF! Hy4, DeepSeek-V4.1-Flash, and Qwen3.8-Flash are also included in the latest offers, covering a wide range of use cases, including coding, advanced reasoning, and multimodal Agent workflows. 🎁 LATEST MODEL DISCOUNTS Here’s a breakdown of the current offers: Hy4: 50% OFF (Limited-Time Offer) DeepSeek-V4.1-Flash: 40% OFF GLM-5.3-Flash: 40% OFF (Upgraded from 30% OFF) Qwen3.8-Flash: 20% OFF Whether you’re focused on software development, complex reasoning tasks, or building multimodal AI Agents, you can choose the models that best fit your workflow. 🚀 INTELLIGENT ROUTING, UNIFIED COMPUTE http:// B.AI offers more than discounted model access. It is building a unified infrastructure for accessing and using multiple AI models. As the Global Settlement Layer for Intelligence, http:// B.AI sits above individual models and beneath AI Agents, connecting different models through intelligent routing and unified computing resources. Developers only need one integration to switch seamlessly between multiple leading flagship models. Every model call is routed, metered, and settled through the platform, helping developers manage usage more efficiently and optimize their AI operating costs. This eliminates the need to build separate integration processes for every model, making it easier to access different AI capabilities through one unified system. 🌐 WHY CHOOSE http:// B.AI? As AI technology continues to evolve, developers need more than access to a single model. They need the flexibility to combine different capabilities for different tasks. Coding: Choose suitable models based on your development requirements. Advanced reasoning: Explore how different models handle complex tasks. Multimodal Agents: Combine model capabilities to support more advanced AI workflows. Cost management: Improve resource efficiency through unified access, intelligent routing, and usage metering. By bringing together multi-model integration, intelligent routing, unified compute, and attractive discounts, http:// B.AI provides developers with a more flexible and efficient AI experience. One integration. Seamless model switching. Intelligent routing. Unified compute. Make every model call more efficient. 🎯 The latest model discounts are now active. Log in to http:// B.AI and explore more AI possibilities! 👉 Get started: https:// chat.b.ai/chat @BAI_AGI @justinsuntron #TRONEcoStar

@
@feisalgro
@feisalgro

Testing Jev by TypeSafe AI on our Google Ads search terms. One big question ("does this fit?") caught 3% of the terms we'd excluded ourselves. Ten simple questions caught 71 to 86%. Fast and cheap is a bonus. The point is better checks.

T
TypeSafe AI@typesafeai
@typesafeai

你上次度假租赁中最昂贵的部分,是你花了四个晚上挑选它的那段时间吗? Jev 把那些晚上还给你

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

出色的工作,展示了校准对实际应用AI的重要性 (我最喜欢的照片附上)

@
@agungmyID
@agungmyID

Pasang https:// typesafe.ai/blog/introduci ng-system-one-models-and-jev … dan https:// github.com/blader/humaniz er … prok..prok..prok jadi apa?

@
@davepoon
@davepoon

Google launched a Decisions API this week, so it went into my Trust Me Bro Phishing test with Jev and the other Decisions models 😄 200 real phishing emails, plus 200 normal emails written to look like them. Same questions and rules for every model. Phishing caught / false alarms: - TypeSafe Jev: 180 / 8 - OpenAI Decisions API: 185 / 3 - Cloudflare Clef: 191 / 7 - Gemini Decisions API: 188 / 1 1 false alarm in 200 is the lowest of any AI model I've tested. It beat Jev on both counts, and the gap is too big to be luck. Clef still catches a few more, but that gap is too small to call. Downsides: about 2x Jev's price, and 709ms per email from Sydney. Around 350ms of that is just the trip to Google's servers.

@
@Mr_Buzzoni
@Mr_Buzzoni

Jev 是 AI 的“互联网”时刻:在使用 Claude Fable 5.2 和 Claude Opus 5.5 的测试中,速度提升高达 193 倍,成本降低 444 倍 Jev 是什么、如何使用它以及如何获得其真正的 100 倍优势,我的 10 页研究: 步骤 1 -> 认识 Jev:LLM 负责写作,代理负责行动,Jev 选择下一步行动 步骤 2 ->

@
@Allen_Schmaltz
@Allen_Schmaltz

A few quick follow-up points (again noting that it's a position paper, so I realize it's meant to be opinionated): 1. Transforms over the output logits of neural networks tend only to be calibrated in a very narrow sense of marginal calibration (cf., approximately conditional)

@
@EnclaveRadar
@EnclaveRadar

🆕 New CVE: CVE-2026-125298 (CVSS 9.8) AsiaInfo Business Support System (BSS) Missing JWT Signature Verification Remote, no login needed. AsiaInfo Business Support System (BSS) 9.2.0 through 9.2.40 does not correctly verify cryptographic signatures for…

@
@PanamaCrypto0
@PanamaCrypto0

In October 2008 Henry Waxman said the CEOs of S&P, Moody's and Fitch, paid over $80 million between them, would testify that virtually no one saw it coming. Then he read Moody's and S&P files. An S&P analyst had written: "it could be structured by cows, and we would rate it." The bill came years later. S&P paid $1.375 billion in 2015 and Moody's $864 million in 2017 to settle with the Justice Department and the states. The hearing that laid out the case costs nothing. In the video below, Waxman reads the documents into the record in a 13-minute opening statement. Listen to what the CEO of Moody's told his own board a year earlier. The House recording has been on the Internet Archive since February 2011. It has 118 downloads. Waxman had done this before. In April 1994 he chaired the hearing where seven tobacco CEOs each said under oath that they did not believe nicotine was addictive. In 2008 he gave his verdict on a new industry: "The story of the credit rating agencies is a story of a colossal failure." Taken together, the documents leave four rules. Ask who pays the grader. The banks selling the bonds paid for the ratings. Revenue at the three firms doubled, from about $3 billion in 2002 to over $6 billion in 2007. Quality has no customer. In October 2007 Moody's CEO Raymond McDaniel told his board that issuers wanted high ratings and investors did not want downgrades. In his words, "ratings quality has surprisingly few friends." Volume drowns out warnings. A month earlier, at an internal town hall, McDaniel said rivals rated everything investment grade in 2004 and 2005, and Moody's warnings went nowhere: "No one cared, because the machine just kept going." Check what the model leaves out. In the full chat, read later in the hearing by Rep. John Yarmuth, the same analyst said the model did not capture half the risk. The SEC had printed the same chat in July 2008 without naming the firm. Those were the CEOs who, Waxman said, would claim virtually no one saw it coming. S&P's settlement alone was about 17 times the $80 million the three had made. Waxman ended his opening with a warning: "The result is that our entire financial system is now at risk, just as the CEO of Moody's predicted a year ago." The full hearing is free on the Internet Archive, posted from the House recording in 2011. It sits under a dry committee title, Part 1 of 2, where few people look. Were the rating agencies wrong to sell grades to the banks that paid for them, or were investors wrong to hand their judgment to a firm the seller paid?

@
@kikifar884
@kikifar884

The partnership aims to merge one of the most recognizable brands in motorcycle culture with @IPXExchange 's financial infrastructure. The model works by enabling the show's community to participate in its success without the creators having to give up ownership of the IP

@
@deeponailabs
@deeponailabs

You can read about system design for weeks. The real round is still a conversation. Someone asks why you picked Redis. Someone asks what breaks at 10x traffic. And you have 45 minutes. SysRound is a voice AI interviewer for system design practice. You talk to Alex the same way you talk to a human interviewer. How a round works: 1. Pick one of 267 questions. Candidates reported them from real onsite interviews. 2. Talk by voice or by text. Draw on a shared whiteboard with notes. 3. Alex, the AI interviewer, answers requirement questions like a product owner. 4. Alex reads the whiteboard while you draw and asks about the hard parts. 5. Alex watches the time and tells you when to move to the next step. A round is 45 or 60 minutes. It has the steps of an onsite round. First come requirements and estimation. Then the API, the data model and a high level design. Then detailed questions on the hard parts, and a summary. What you get at the end: - A score for each of seven skills, compared with the level you target - Quotes of what you said or drew, next to each score - A model answer that starts from your own notes - A minute by minute replay of the whiteboard - A practice plan for your weakest skills During the beta, you get 2 free rounds of 45 minutes. After that, rounds come in packs. There is no subscription. The AI interviewer and the feedback can be wrong. Use it for practice, not for hiring. How do you practice the talking part of system design today?

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

电子表格用户体验的疯狂重构: https:// practicaldata.dev 🤯 使用jev来语义化高亮单元格 (我们花了好一阵子试图搞出炫酷的Excel演示,这个要酷炫得多)

@
@The_profray
@The_profray

What if you did not need to hire one team to improve an AI system? What if many independent teams could compete to build something better? That is the question claims is asking while also providing the solution to that answer. @DeSciClaims (SN111) turns scientific papers into structured records but getting this right is not easy. A miner must identify the claims in a paper and link them to the right evidence. Poor work should not pass just because it sounds convincing. This is where the competition comes in and you might just understand it better. Miners submit their work, and blind AI judges review the claims, the results help decide which submissions meet the standard. The aim is to reward better work, not just more submissions. The new competition is designed to use one tested pipeline in production instead of running about ten versions for every paper. Claims estimates this could mean 10 – 100x more papers per dollar of emissions. That is a cost model estimate, not a result already achieved.

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

街上写着系统2,床上却是系统1? @swyx 你是在暗示我所想的那样吗?👀

T
TypeSafe AI@typesafeai
@typesafeai

哦,所以现在每个人都在轻松地“氛围编码”自己的应用?就差一个垃圾大炮,就能得到生活中想要的一切?证明给我看! 今天就让你的梦想超速起飞,变成现实

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

非常糟糕!我们对第一波杰文斯悖论化智能的猜测将是“暗数据” 为什么我们不把智能扔到我们已经拥有的海量数据上呢?太昂贵 + 在 map-reduce 中太不可靠 但显然杰文斯改变了这两个问题 (:

T
TypeSafe AI@typesafeai
@typesafeai

你可以将你的语义逻辑偷偷塞进字符串的细则中、嵌入向量中,以及所有你想要的复杂机器学习模型里,但数据分布总是会比你能想到的任何聪明点子都要荒谬得多 所以: • 明确你需要知道什么 • 明确地说明,对你数据的直接解读就足以让你知道所需信息

T
TypeSafe AI@typesafeai
@typesafeai

所以我们听说你喜欢决策模型 但是你如何用模型对复杂的数据做出可靠的决策呢? 你做不到:可靠性需要代码,因此可靠的决策是通过在代码中组合可靠的判断来实现的 Jev 做出可靠的判断

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

一篇绝对美丽且信息丰富的博客文章!我们在jev中看到的最大反模式之一,就是直接把它插进去来做重大决策,而不是让代码做它最擅长的事情,而把jev只用于那些需要判断的部分🤌 这才是我们改造软件的方式!

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

在无监督学习播客上度过了一段超级有趣又火辣的时光!!!🌶️

T
TypeSafe AI@typesafeai
@typesafeai

快速构建它。构建正确的东西。 HYPERSHIP 在 @ProductHunt 上,本周五。

T
TypeSafe AI@typesafeai
@typesafeai

Jev 在哪里用?Jev 无处不在! (错失的机会) 但说真的,我们才刚刚开始探索可能的智能软件空间,你们却在想得太小了?

T
TypeSafe AI@typesafeai
@typesafeai

你闻得到那股味道吗? 那是成千上万的开发者在周五像上帝所愿的那样推送代码到生产环境,多亏了 @ProductHunt 我们将在11月6日与你再见

T
TypeSafe AI@typesafeai
@typesafeai

介绍 Jev @typesafeai 驱动的人力搜索 $0.0024/线索。不再有 $2,000 的订阅费。 您的代理现在可以: > 级联遍历 Apollo、Clay、People Data Labs、Hunter、ContactOut、Ploid > 通过 Jev 处理 1,000 个线索 > 仅为 Jev 保留的线索支付数据丰富费用 开源。无订阅。0% 加价。

T
TypeSafe AI@typesafeai
@typesafeai

人们想要 swag。我们听到了。所以理所当然:免费冰激凌代替。 @Judgmentlabs 用免费冰激凌庆祝,我们将供应 Jev-Jam 口味的冰激凌勺! 235 Pine St, San Francisco 星期二,10 月 27 日,上午 11 点至下午 1:15

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

尽管时机不幸,我们在 @typesafeai 非常认真对待披露事项 大约在本周五晚上6:49,我们了解到所有客户noul输出的巨大未经授权泄露 为了最大限度地提高透明度,这里是链接: https:// gist.github.com/diogo149/81003 7374943dae22ff251941346b4e8 …

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

官方回应:如何获得jev周边商品见下文 (开玩笑的,我们正在努力搞定 🙈)

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

激动人心 :D 我们爱 Dagger (尽管我们几乎不认识她)

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

大多数AI公司:令牌工厂 我们的数据团队:托尔金工厂

T
TypeSafe AI@typesafeai
@typesafeai

定制代理,配备定制知识库、定制指令、定制工具集,当然,还有与Jev的定制路由 让你的goon成为你独有的!

T
TypeSafe AI@typesafeai
@typesafeai

对于信任问题,有时你需要心理治疗 但你的代理需要一个可靠的代码防火墙,由 Jev 提供支持

T
TypeSafe AI@typesafeai
@typesafeai

Jev 为 @UiPath 工作流程带来可靠的智能判断,大约便宜和快一个数量级 LLM-as-a-judge-no-more

T
TypeSafe AI@typesafeai
@typesafeai

老实说,Cookie 横幅是个糟糕的主意 接下来应该自动化掉什么?

T
TypeSafe AI@typesafeai
@typesafeai

@Corderabbitai 的 Jev 黑客马拉松所展现的创造力和生产力简直是疯了。 他们的视频里有许多精彩的小片段!

T
TypeSafe AI@typesafeai
@typesafeai

在这艰难时刻,我们能为您提供一个 TypeGlizzy 吗?

@
@0x_rody
@0x_rody

Jev founder just packed everything published on Jev Engineering into a 1-page PDF One formula stops your agent from paying a frontier model to say yes or no: Decision Layer = LLM Writes + Jev Decides + Code Acts here's the full setup: step 1 -> let the LLM write the plan, then list every decision point in your loop: route, gate, retry, escalate step 2 -> turn each one into a question: Choice to pick a route, Score to rank priority, Noul for yes/no step 3 -> send the state and all your questions to Jev in one call, you get probabilities back step 4 -> put the thresholds in your code: if Noul says a Bash command is ≥0.7 destructive, block it, at ≥0.35 ask first step 5 -> if Jev isn't confident enough, send it to Claude and don't apply it automatically TypeSafe charges $0.042 per million input tokens, output is free full breakdown in the video 👇

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

说得太好了!关于如何真正让模型变得更智能,以及我们为什么招聘前端/全栈工程师

T
TypeSafe AI@typesafeai
@typesafeai

嗯,也许一个基准测试还行,如果它让你饿了的话

T
TypeSafe AI@typesafeai
@typesafeai

谁来基准测试基准测试者? 一个好的评估是: - 短暂的。它捕捉了时间和能力的某一瞬间。 - 基于事实的。它直接衡量您关心的任务和数据上的性能。 - 具体的。它针对期望的结果进行操作,而不是某种替代指标。 记笔记!

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

兄弟们,请救我脱离 @justKDeng (她想让我变得无聊)

T
TypeSafe AI@typesafeai
@typesafeai

我们的首席执行官 @CompleteSkeptic 讲述了爱因斯坦博士和憨豆先生的轶事 > 如果人工智能 95% 的时间像爱因斯坦那样回应,5% 的时间像憨豆先生那样,你要如何用它来自动化任何事情? @mattturck 向我们介绍了 Jev 疯狂的第一个月

T
TypeSafe AI@typesafeai
@typesafeai

没有我们的饮料,我们什么都不是 h/t johnhq from our discord

T
TypeSafe AI@typesafeai
@typesafeai

没有我们的饮料,我们什么都不是 h/t johnhq from our discord

T
TypeSafe AI@typesafeai
@typesafeai

有趣的是,这些动态设计、结构化工作流程和知识任务解决方案突然间出现的速度越来越快,仿佛自动化它们变得容易得多。 不知道发生了什么 🤔 干得漂亮 @Lovable

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

不过别看这场打斗怎么结束 https:// x.com/hellokillian/s /hellokillian/status/2110761114065670607 …

T
TypeSafe AI@typesafeai
@typesafeai

我们在 10 月 14 日太平洋时间下午 1:14 宣布了 Jev 1.14 🔥 一个全新且改进的模型,连同我们梦想家们的实验性功能: • 图像 • 一种新的盒子类型 • 更长的上下文 惊喜,Jev 已经能够看到你了!👁️ 但我们对生产级智能的标准相当高,所以建造者们请注意,这些是预览功能。 自行进入 ~*~

D
Diogo Almeida@CompleteSkeptic
@CompleteSkeptic

快乐的 JevDay! 但在公布之前,我们得先聊聊哲学…… 早期采用者 / 捣鼓者 / 梦想家的需求,与极其稳定的基础设施所需的需求截然不同。 对于 Jev 版本,我们倾向于后者,但这个版本是为前者准备的——带有一个关卡,让人们可以选择进入危险区!

T
TypeSafe AI@typesafeai
@typesafeai

这是什么? Jev 正在进化! 请在下午 1:14 收看我们与 @CompleteSkeptic 和 @allietheicon 的市政大厅会议,以了解更多

T
TypeSafe AI@typesafeai
@typesafeai

这是什么? Jev 正在进化! 请在下午 1:14 收看我们与 @CompleteSkeptic 和 @allietheicon 的市政大厅会议,以了解更多

@
@thedelost
@thedelost

Anthropic 刚刚发布了 Claude Code 的最佳技巧: Opus 5.5 规划 → Haiku 5.5 阅读仓库 → Sonnet 5.5 编写 diff → Fable 5.2 签发确认 1 分钟 + 一行 frontmatter = 你的舰队停止运行 Opus 首先,理解它的工作原理: • 每个子代理都有自己的上下文窗口,只有其摘要会返回,因此它读取的 214 个文件永远不会触及你的主会话 • .claude/agents/ 中的 model 字段决定每个代理运行在哪个模型上:haiku、sonnet、opus、fable,或完整的模型 ID。文件中的其他内容不变 • Claude Code 默认同时运行多达 20 个子代理,它们可以嵌套三层深。这意味着你要为大量回合支付前沿费率 以下是设置方法: 1 → 在仓库中创建 .claude/agents/,或者如果你想在每个项目中使用,创建 ~/.claude/agents/ 2 → 每个代理一个 markdown 文件。只有名称和描述是必需的,正文成为其系统提示 3 → 对于仅读取的一切,设置 model: haiku。搜索、grep、文档、标签、摘要。这些都不需要前沿模型 4 → 对于编写 diff 的那个,保留 model: sonnet,对于审查的那个,设置 model: fable。该字段的有效值包括 Fable 5.2,因此你的审查代理会在其上运行,而不影响其他任何内容 5 → 没人告诉你的:省略 model 字段,子代理会悄无声息地继承你的主模型。你的 grep 代理一直都在 Opus 上运行。设置 CLAUDE_CODE_SUBAGENT_MODEL=haiku,它会捕获你忘记的每个文件 还有一点:保持描述简短。组合超过 15,000 个 token 时,Claude Code 在启动时会警告你,而且你每次回合都要为所有这些付费。 我会在下面留下完整的提示和 Anthropic 关于子代理的官方文档 然后阅读下面的文章 ↓

Hacker News