Austin Walters

Turns out I can de-anonymize users on #HackerNews (already turned this system off): hnprofile.com With great power comes great responsibility, I can only imaging what @Snowden saw... news.ycombinator.com/item?id=179442…

@HedgieMarkets @GlenWilsonIA Inference costs are already an order (or more) magnitude lower if you host yourself.

@GlenWilsonIA 🦔In my opinion, replacing human labor at scale requires inference costs to collapse by an order of magnitude, and the trend over the last six months runs in the opposite direction. The 18-month predictions assume cost trajectories that no longer match the price sheets.

🦔Microsoft canceled its internal Claude Code licenses this week after token-based billing made the cost untenable, even for a company with effectively infinite cloud resources. Uber's CTO sent an internal memo warning the company burned through its entire 2026 AI budget in just four months. American AI software prices have jumped 20% to 37%, and GitHub (owned by Microsoft) is dropping flat-rate plans for usage-based billing across its products. My Take The AI subsidy era is ending in real time. The same company that put $13 billion into OpenAI and built the Azure infrastructure powering most of Anthropic's compute just looked at the bill from a competitor's coding tool and decided it was not worth paying. That is not a productivity failure on Anthropic's end. Token-based pricing is forcing every enterprise customer to confront the actual cost of running these models at scale, and the number turns out to be far higher than the flat-rate experiments suggested. This ties directly to my Gemini Flash post yesterday. Anthropic, OpenAI, and Google all raised effective prices in the last six months. Enterprises that built workflows assuming AI costs would keep falling are now watching annual budgets evaporate in months. Two outcomes look likely from here. Either enterprises scale back AI usage to fit budgets, which slows the revenue ramp the labs need to justify their valuations ahead of IPOs, or the labs cut prices and absorb the losses, which makes the unit economics worse at exactly the wrong moment. Both paths land in the same place, the numbers stop working, and somebody has to take the writedown. Hedgie🤗

🚀 Meet Qwen3.6-27B, our latest dense, open-source model, packing flagship-level coding power! Yes, 27B, and Qwen3.6-27B punches way above its weight. 👇 What's new: 🧠 Outstanding agentic coding — surpasses Qwen3.5-397B-A17B across all major coding benchmarks 💡 Strong reasoning across text & multimodal tasks 🔄 Supports thinking & non-thinking modes ✅ Apache 2.0 — fully open, fully yours Smaller model. Bigger results. Community's favorite. ❤️ We can't wait to see what you build with Qwen3.6-27B! 👀 🔗👇 Blog: qwen.ai/blog?id=qwen3.… Qwen Studio: chat.qwen.ai/?models=qwen3.… Github: github.com/QwenLM/Qwen3.6 Hugging Face: huggingface.co/Qwen/Qwen3.6-2… huggingface.co/Qwen/Qwen3.6-2… ModelScope: modelscope.cn/models/Qwen/Qw… modelscope.cn/models/Qwen/Qw…

Bureaucracy for AI agents

TL;DR: Top hacker calls Anthropic’s bluff.

BREAKING: While a new War for Oil erupts in the Middle East A Physics Paper just quietly dropped TODAY that will eventually make Oil, and the entire current Energy Industry, irrelevant. Ushering in the era of Zero-Point Energy @EagleworksSonny Here is the breakthrough🧵

It occurs to me that we're seeing a pattern here.⬇️ The Iranians are probably baiting our air and missile defense radars in the Middle East into illuminating and giving up their exact positions with ballistic missiles and then killing them with drones. Their cruise drones are by all indications quite precise and effective enough at sliding through fighter sweeps and the very limited SHORAD coverage that exists, likely because the Iranians got a dress rehearsal last year and the Russians have shared industrial amounts of combat data from Ukraine with them. We've lost a terrifying proportion of the total number of THAAD radars in existence just in the last week, not to mention some ultra-heavy fixed radar installations and an unknown number of Patriot radars. We're already seeing the results of this. Missile warning times in Israel have already decreased to a matter of seconds. Missile warnings in the Gulf States are sporadic to nonexistent. Replacement systems are being frantically flown into the Middle East from elsewhere. And the Iranians are getting essentially the same number of missiles through and continuing to strike important targets despite significantly fewer launches - suggesting they're metering their launch campaign by effects on target to sustain a long war. This also, by the way, suggests that we got head-faked by Iranian strike doctrine. We've been laser-focused on countering the flashy high end of their strike system - the ballistic missile force, while the Iranians didn't use their drones very effectively in previous rounds. During the Twelve Day War the trickle of Shaheds the Iranians flew off was trivially interdicted by fighter sweeps, which may have lulled US and Israeli war planners into a false sense of security about the true effectiveness of these weapons. This may have been precisely the impression the Iranians wanted us to have of what we know, from extensive combat use in Ukraine, is a pretty formidable weapon. The Iranians may actually have taken a pretty serious lesson from Ukraine (that flashy weapons draw fire) and shifted to considering their drone force - easily produced in gigantic quantities, quiet to launch, easy to store, frustratingly difficult to intercept, and lethal and precise enough to disable key nodes - to be their primary striking arm.

@mark_l_watson How’s the performance of qwen3.5 35b compared to the prior model (qwen 3 30b)?

I have been experimenting with Liquid's new lfm2 model that runs fast on a 32G Mac mini, and so far, handles all of my test tool use cases. I really like qwen3.5:35b (a lot) but lfm2 runs much faster on my system.

The 70% Life Extension Nobody Is Talking About Two cheap drugs. One university lab. A 73% increase in remaining lifespan for elderly mice. The longevity breakthrough that should be front-page news and why decentralized science may be the only force that can bring it to your medicine cabinet. Oxytocin and Inhibitor Extend Mouse Lifespan
 Combining oxytocin and Alk5 inhibitor extended elderly mouse lifespan by 73%. Inexpensive drugs, no gene therapy. Major longevity paper. 
Why Important: Demonstrates affordable pharmacological approaches to longevity, bridging natural hormones with targeted inhibition for human applications. doi.org/10.18632/aging…

@0x49fa98 @Aristos_Revenge More to do with hardware and frameworks, LSTMs are effectively able to get to the same accuracy levels as transformers. Transformers have a couple of advantages, but in terms of accuracy, there are models with 1970s math that would work as well as at least gpt-3.5

You change one word on a loan application: the religion. The LLM rejects it. Change it back? Approved. The model never mentions religion. It just frames the same debt ratio differently to justify opposite decisions. We built a pipeline to find these hidden biases 🧵1/13

Yeah wonder how else they’re using this tech in coordination with police departments…

People ask me all the time about compelling use cases of AI. Here’s a good one. Millions of dogs go missing in the U.S. every year—and options for finding them are often painfully limited. Our Ring team saw an opportunity to use our community and technology to help, so they built Search Party. When a pet owner posts about a lost dog in the Ring app, nearby participating outdoor Ring cameras in the neighborhood begin looking for potential matches. If yours spots what might be the missing dog, it lets you know. You see the photo alongside footage from your camera, then can choose to share the video with the pet’s owner. The AI is trained on tens of thousands of dog videos so it can recognize different breeds, sizes, fur patterns, body features, unique marks, shape, and color. And privacy stays in your control—you decide each time whether to help. The impact is energizing. Search Party has helped bring home 99 dogs in just 90 days—more than a dog a day since launching three months ago. Ring customer Kylee was blown away by Search Party after her dog Nyx was found by a neighbor’s camera just 15 minutes after slipping through a tiny hole he’d dug under her backyard fence. When a Ring customer and military veteran named Kurt realized his service dog was missing after jumping his fence, he worried he might have lost her for good. He quickly initiated a Search Party in the Ring app asking neighbors to help locate her. Later that day, he got the notification he was hoping for…Lainey was found. Chris, a Ring camera owner, helped reunite another lost dog with its family after getting an app alert that said, “Your camera may have spotted a missing dog,” flagging footage he wouldn't have otherwise noticed. And the list of stories like these keeps growing. Now we’ve expanded this feature so that anyone in the U.S. can start a Search Party through the Ring app, even without a Ring camera (lost pets are one of the most common posts in the Ring Neighbors app—over 1M last year alone). With roughly 90 million dogs in the U.S., think this is gonna matter for a lot of families. Good example of real-world impact, and proud of what the Ring team has built here. aboutamazon.com/news/devices/r…

Kind of old news, but qwen3-coder-next with Claude Code is really impressive for a low inference cost model.

@96Stats @mark_l_watson Market forces will drive the US, China is centrally managed. You don’t think every company is implanting AI in the US?

Very surprising read by the WSJ!! As someone who works in AI in China, i fully agree too. The US is chasing AGI making bigger models, more chips, higher reasoning ceilings. Pour in billions, build godlike systems etc etc then hope business value comes in Yet here in China, it's completely different, even though they had all those US chip sanctions. The government have basically told Deepseek to focus on implementing their AI into society rather than making a godlike model.. healthcare, education, government..things to benefit everyday people rather than just a small number. That's what AGI is for China. In just the past couple months AI has been used in sooo many areas, especially where i live here in Xinjiang, they've used AI in steel plants, to coal mines 700m underground, to run factories unmanned etc. Also, i'm working with a hospital in Urumqi where we use AI to cut cancer diagnosis time from minutes to seconds (saves a lot of time longrun with hundreds of patients a day btw)

BREAKING 🚨 TESLA LOCKS DOWN THE "SECRET RECIPE" FOR ITS DRY ELECTRODE MANUFACTURING 🔒 For years, the battery industry believed that mass-producing dry electrodes was impossible, a lab trick that simply couldn't scale. Published on January 29, 2026, patent application US20260031317A1 proves them wrong and reveals the next, ruthless phase in Tesla's intellectual property strategy. If the previous patent was about owning the car, this one is about owning the factory. This filing serves as the definitive "cookbook" for the holy grail of battery manufacturing. While Tesla has already secured the rights to the superior performance of the battery, this continuation protects the method. By patenting the exact order of operations and physical constraints required to ditch toxic solvents, Tesla is effectively copyrighting the "kitchen" so that no one else can bake the same cake. This ensures that even if competitors figure out what makes the dry electrode work, they will be legally barred from using the most efficient way to make it. To understand why this legal firewall is so necessary, we have to look at the specific engineering trap that Tesla is trying to prevent competitors from exploiting. 🧩 The problem: Copying the result, evading the method The transition from "wet" to "dry" manufacturing is notoriously difficult because of a cruel physical trade-off: to make dry powder stick together into a solid sheet, you typically need to apply high-shear force or add large amounts of polymer "glue". Both are bad. High shear crushes the delicate battery crystals (killing lifespan), while excess glue wastes space (killing range). Tesla has solved this by developing a "Goldilocks" zone, a gentle mixing process that activates the binder without destroying the particles. However, this creates a legal vulnerability. In the world of patents, securing the "end product" (a high-efficiency battery) is a massive win, but it leaves a loophole. Competitors could theoretically try to achieve similar battery performance using a slightly different, less efficient, or messier process to skirt the patent rules. If Tesla only protects the final battery, rival manufacturers could reverse-engineer the specifications while claiming their production line is "different enough" to avoid infringement. To truly secure its competitive advantage, Tesla needs to protect the unique, low-cost "kitchen" where the battery is made, not just the "cake" that comes out of the oven. To close this specific loophole, the new filing moves to secure the manufacturing process itself. 💡 Tesla’s solution: The "method" is the moat Tesla’s solution, detailed in this continuation, shifts the legal focus from the device to the method of fabrication. This effectively means Tesla is moving from protecting the final battery product to patenting the specific recipe and cooking steps used to make it. The key innovation here is not just that the electrode works well, but that it is manufactured using a specific, counter-intuitive sequence. The patent application seeks to protect a method that involves nondestructively mixing active materials with porous carbon first. These active materials are the primary lithium compounds that actually store the energy, while the porous carbon acts as a conductive additive that functions like a microscopic electrical grid. The process uses nondestructive mixing, which is a gentle blending technique that mixes the ingredients without crushing them, much like folding ingredients into a cake batter to keep it airy. Only after this initial blend is complete does the method involve adding the dry binder to create the final film. This dry binder is a polymer adhesive that serves as the structural glue to hold the powder mixture together in a solid sheet. By legally defining this specific order of operations, specifically mixing the dry energy-storing ingredients before introducing the glue, Tesla is fencing off the most logical and efficient way to produce dry electrodes. This prevents competitors from adopting Tesla’s streamlined manufacturing flow, forcing them into less efficient, more complex, or more expensive production methods. But the "method" is only half the story; the other half relies on the specific physical characteristics of the ingredients themselves. 🔬 The innovation: Large particles and "gentle" manufacturing This filing doubles down on a specific physical constraint regarding the size of the particles used in the battery. The patent explicitly claims protection for using active material particles that are at least 10 microns in size. For context, ten microns is roughly one-tenth the width of a human hair. This is significant because traditional battery manufacturing often relies on pulverizing materials into fine dust to make them fit into a wet slurry, which is essentially a muddy paste created by mixing powders with liquid solvents. Tesla has discovered that by keeping the particles larger and pristine, they can use significantly less binder. Specifically, they use less than 2% by weight of this binding glue. The patent describes a process where these larger particles serve as the structural bricks of the electrode wall, while the PTFE binder acts as the minimal mortar. PTFE (polytetrafluoroethylene) is the same polymer found in non-stick cookware. To achieve this structure without cracking the large particles, the method specifies using acoustic or low-speed blade mixers running at a crawl of 10 to 40 meters per minute. Acoustic mixers use sound energy to vibrate and blend materials without direct contact, while blade mixers gently fold the powder like a slow-moving dough hook. This nondestructive approach is now a core part of the claim, ensuring that the method itself is recognized as a unique invention because it preserves the original quality of the materials. With the physical method established, Tesla tightens the noose further by adding strict chemical rules that make the patent nearly impossible to sidestep. 📝 The fine print: Three critical constraints To truly lock out competitors, this continuation filing adds three hyper-specific "fences" around the manufacturing process that move beyond general concepts to define the exact chemical and physical limits of Tesla's technology. First, the patent imposes a strict "Single Binder" rule. While many battery manufacturers use a cocktail or complex mixture of glues to balance adhesion and flexibility, often mixing PTFE with other polymers like PVDF (polyvinylidene fluoride) or CMC (carboxymethyl cellulose), Tesla’s filing explicitly prohibits this. The text specifies that the binder "consists essentially of a single dry fibrillizable binder". A fibrillizable binder is a material capable of forming a microscopic web of fibers when mechanically stressed. This forces the recipe to rely 100% on the mechanical fibrillation of PTFE, a process that physically stretches the binder particles into thread-like networks rather than relying on chemical stickiness. It asserts that their process is so refined they don't need the chemical crutch of secondary glues. Second, Tesla places a hard legal ceiling on conductive carbon. This carbon serves as an electrical pathway but acts as a "dead weight" filler because it does not store any power itself. The patent caps this material at "at most 8 wt%", meaning it can comprise no more than eight percent of the total weight of the electrode. While carbon is essential for electricity to flow, it stores no energy. Competitors might try to make a dry electrode work by dumping in 15-20% carbon to compensate for poor connectivity, but that results in a mediocre battery with less room for active ingredients. By setting this limit, Tesla protects the high-performance version where filler is kept to a bare minimum to maximize energy density, which is the amount of energy stored relative to the battery's size. Finally, the filing reveals a "Hero" configuration that proves this process isn't just for lower-end standard batteries. It details a specific formula using 98% NMC 811. NMC 811 refers to a lithium nickel manganese cobalt oxide chemistry rich in nickel, which is difficult to handle but offers superior range. This formula combines that high-performance material with just 1.25% PTFE binder and 0.75% total carbon. Achieving a stable film with such a high load of active material proves this dry process is ready for Tesla's most demanding vehicles, effectively turning the electrode into a nearly solid block of energy. Once this highly specific mixture is prepared, the final step of the process seals the advantage by defining the speed of production. ⚡ The "3-pass" efficiency A critical detail in this continuation is the speed of formation. The filing highlights that this specific recipe allows the dry powder to be turned into a self-supporting sheet, meaning the film is structurally sound enough to be handled like a roll of fabric without crumbling or needing a supporting metal foil. This result is achieved after passing through a process called calendering, which involves feeding the material through a series of heavy steel rollers that press it flat, much like a pasta machine flattening dough. The patent specifies this happens at most three times. In manufacturing, fewer passes equals higher speed. By claiming a process that creates a sturdy film in just three steps, Tesla is effectively patenting the velocity of its production line. This velocity determines the overall factory throughput, or the volume of finished product made per hour. A competitor trying to replicate this might need 10 or 20 passes to get a stable film. This requirement would force them to run the material back and forth repeatedly, making their factories slower and more expensive to run than Tesla's. This combination of legal, chemical, and manufacturing speed constraints lays the foundation for Tesla’s dominance in the next decade. 🚀 How this continuation contributes to Tesla’s now and future First, it blocks competitors from "fast-following" the 4680 production method. While other automakers can buy good batteries, this patent prevents them from building factories that operate like Tesla’s. By protecting the specific "mix-then-bind" sequence and the "low-binder" recipe, Tesla ensures that its Gigafactories remain unique. Competitors cannot simply buy the same mixing equipment and run the same recipe without risking patent infringement. Second, it secures the economics of "cheap" raw materials. By specifically patenting the use of larger (greater than 10 microns) particles, Tesla is validating a cheaper supply chain. Smaller, highly processed particles cost more. This patent confirms that Tesla’s process is optimized for standard, "bulk" grade materials. Protecting this capability ensures Tesla retains a cost margin advantage, as they can turn cheaper, commoditized inputs into premium performance outputs. Third, it creates a legal "thicket" around dry electrode tech. This filing is a classic "picket fence" strategy. The parent patent protects the battery efficiency (90-94%). This child patent protects the binder loading (less than 2%) and the particle size. Future filings will likely protect the machinery. This layering makes it nearly impossible for a competitor to design a dry electrode without tripping over at least one of Tesla’s patents. Finally, it validates the "micro-factory" concept. The emphasis on creating a "free-standing" film without a metal foil backing is crucial. It means the electrode film can be made in one machine and rolled up, then applied to foil later. This decouples the manufacturing steps, allowing Tesla to fit production lines into smaller, non-linear spaces, which is essential for the tight footprints of future factory expansions or retrofitting existing lines.

Important context is that many courts have held that forcing people to unlock devices with biometrics does not violate the Fifth Amendment privilege against self-incrimination.

If you are wondering why I have said so little about Minnesota, I can all but assure you that many of you would find what I would have to say pretty upsetting. Let’s start here: I think we are moving away from rule of law and towards executing people in the streets under FAFO, and that there are many in the streets who are revolutionaries *currently* cosplaying at Non-Violent Civil Disobedience. There. That should be enough to piss off everyone on a team giddly dragging us towards armed civil conflict. You know who you are. I don’t want armed civil political conflict. I just don’t. In the U.S.? Have we lost our minds.

I shipped more code in the last 3 weeks than the decade before. The top AI models / agentic systems right now are an entirely different thing to what people used until the beginning of December.

For my next prediction: AI training via massive data collection has probably already reached its top potential. The smart people say we need to build virtual digital worlds so AI can learn the way humans learn, by living and interacting in the artificial reality. Human-like characters will inhabit the artificial reality, so robots get the most like-us training experience. The artificial creatures will look and act like people, and be programmed to believe they are base reality. The number of artificial realities will quickly surpass the number of real ones, which we think is one. Put it all together and it will become increasingly obvious that what we think is our base reality is PROBABLY a simulation. This is the year.

@DavidSHolz We have entered the Singularity