Im starting to suspect the push to make AI models more efficient is costing us more than we’re being told. This goes for both Anthropic and Open-AI I genuinely trusted GPT-5.5 with multi-step work more than I trust Astra. Getting a similar feeling with newer Opus releases. Everyone’s posting benchmarks and impressive one-shots, but that’s not how I use these things. I’m working on existing projects where following instructions and not breaking unrelated stuff actually matters to the reliability of prod. models reach for computer use when a simple CLI command would’ve done the job. Other times, something I explicitly asked for just never gets implemented. like, the model doesnt hit a blocker, it just forgets, when later caught, it says, my bad i should have made it more clear before.. this is why i cant implement this actually My suspicion is that they’re optimizing serving costs harder than we realize. Smaller models/Fewer active parameters/Too much RL or smth? This affects the vibecoders more than it affects me, i thankfully know what is wrong with the code rather than codex telling me about what it is. quantization is another possibility. Blackwell already supports NVFP4, and Vera Rubin is built to accelerate it further. I’m not saying "4-bit makes models stupid" or claiming I know what precision these providers use. But I’d like to see whether the serving optimizations preserve reliability across messy, multi-step work and not just the average score on an evaluation. I get why they want cheaper inference. I want cheaper inference too. I just don’t want an upgrade that needs more supervision than the model it replaces because then that would mean more tokens wasted and the efficiency wouldn't really matter. the models have gotten great at one shots, they truely have. but that is it, they are great at one-shots.. not actual multi-step prod work, this is why you are seeing a surge of bugs literally everywhere. it is not a coincidence. every developer at big tech agrees with this. when it comes to /goal mode.. i think its the most stupidest thing there is, sure it keeps the model aligned to the goal. but when the model actually gets stuck at something, it really years a developer input, i have seen some thinking traces before that prove this point, the model really thinks about stopping and asking the developer what to do next, i think models cant really think multi-step and branched thinking. they think linearly and they cant really think of two things at the same time. and most lazy ass developers set goals like finish the implementation.. this is genuinely stupid. Vibecode frontend completely, yes.. do that. but please do not vibecode backend, trust me its very hard to maintain, when you try to implement a new feature, it ends up breaking stuff that you have to fix again, this is where 5.5 was good at, when i prompted it to implemnet somethign and then do actual regression testing it would remember stuff a lot better and did deeper regression testing. look I don’t know their internals, so I’m not pretending any of that is proven. But more efficient for whom exactly? we’re losing reliability along the way and calling the whole thing an upgrade. If I have to repeat instructions, babysit every action and fix the damage, the cost per token going down, in most cases means more token consumption. i think most people, dont really know how to tell a difference between models at this point, give them a sort of blind test between grok and fable and they are not really able to tell which is which, this is because all the models have crossed the boundry that is what i call, "Basic human intelligence" and hence the models are, in most cases going to suggest things that are novel to an average human. And i would argue that Sol, astra are more Sycophant about stuff than any of the claude models have ever been. astra has never said that we should do this rather than this, whereas claude always has. it feels like claude doesnt care about me being sad with what it answers, perhaps it could shatter my expectations from a particular idea, but that would mean that the idea was stupid in the first place. astra and sol never want that to be the case, they care more about pleasing the user. try this, say somethign stupid and tell the models to evaluate it, use fable and astra side by siode, tell both models to call out stupid when you (user) are stupid and see which model hesitates the most and which gives genuine advice. i think genuine advice matters a lot, its liek claude treats me like a gigachad senior developer. i would really like to know your thoughts on this, please share them below. in purely technical details if possible.

Every piece of software is having 2-5x bugs right now, and vibecoding is to blame for a lot of it. enginners have stopped reading the code they are shipping to production

All the developers who like Jev are not developers, they are vibecoders Yet to see someone who actually knows coding to say that they use Jev for something because I have not been able to find a use-case for it at all. Please prove me wrong. Imo its group-think twitter hype

nypost.com/2026/09/16/lifest… Must. Sequence. Surfactant. Bugs.

A drug that has been sitting in the pharmacy since 1999 roughly doubled survival in the deadliest brain tumor children get. It is not doing it by stopping seizures. doi.org/10.1038/s41591-026-0… — Barron & Monje, Nature Medicine

LinkedIn is discovering the fly 😭🙏

I absolutely love that when someone says life sciences whay they really mean is drug discovery. THERE IS MORE TO BIOTECH THAN PHARMA FFS.

Today in @Nature, we report a new approach to studying human brain development and modeling disease in vivo. Human neural #assembloids and #organoids have opened new ways to study development in vitro, but they also have important limitations. Transplantation can provide a more physiological environment, yet human cortical cells then develop within a rodent brain that matures much faster, constraining their growth and integration. To address some of these limitations and obtain more advanced functional readouts relevant to disease biology, we generated #apallial mice, in which ~98% of the cortex and hippocampus fail to form, creating space for transplanted human cortical organoids to grow and integrate extensively. We call these #xenocortical mice (XCX). Nearly 7 years in the making. Congratulations to an extraordinary team and wonderful collaborators! Link to the article below 👇

They deleted most of a mouse's cortex before it was born, filled the empty skull with human neurons, and the mouse walked, learned and remembered doi.org/10.1038/s41586-026-1…

Looking at some old pancreas slides

This is interesting technology, but it is not a replacement for generative LLMs, and comparing it directly with frontier LLMs creates the wrong impression. Jev is essentially a natural-language decision model. You give it context, questions and a predefined set of possible answers; it returns probabilities over those answers very quickly and cheaply, with many judgments evaluated in parallel. it is an excellent classifier and ranking engine for routing, risk scoring, checking agent outputs etc however, It cannot write, code, explain its answer or propose an action outside the options supplied. A frontier LLM can invent the menu; Jev is optimized to choose from the menu. Those are different jobs. So claims like 200× faster than frontier models and 400 times cheaper is stupid Also, We already had systems liek these and they are easy to build these days: Rule engines, decision trees, task-specific classifiers already handle these workflows. and they do it for much lower cost (basically free) and they are easy to audit and change as well Jev’s actual claim is that it can perform arbitrary natural-language-defined judgments without task-specific training, evaluate many of them in parallel and do so at classifier-like speed and cost. The claim that it cannot hallucinate also cant be true.. Jev may always return a structurally valid option, but it can still return the wrong valid option. Type safety prevents malformed outputs but it does not guarantee correct judgments.. it cant do that. when Jev reports 90% probability, is it actually correct around 90% of the time on independent, unfamiliar data? If that can be demonstrated reliably, Jev could serve as a cheap, always-on semantic reviewer wgole job is flagging agent actions that appear unauthorized, destructive or inconsistent with policy before they are executed. Sort of an agent-actiosn firewall But this is not a missing layer in modern agent infrastructure for eg: Codex already combine contextual model-based review with tool schemas, deterministic rules, OS-enforced sandboxing, scoped permissions etc. Those controls are stronger wherever risk can be expressed precisely because they enforce boundaries rather than merely predict danger. So basically RLCD has not demonstrated enough unique value to justify paying for it at all

Solar panels should be on all roofs, huge advocate for photovoltaic glass too, particularly on skyscrapers.

Can anyone form @supabase help me, for some reason I am logged out of my account and my TOTP has disappeared from my Google Authenticator Is there a way I can get my account back? I can prove ownership and i have all the billing invoices

Incredible performance today from the Red Arrows, always look forwards to their flyovers every year

I want to talk about this Navier Stokes result for a minute because I see alot of people sharing it and not nearly enough people explaining it. A share for this one would be appreciated. The Navier-Stokes equations are just Newton's second law of motion, F=ma, adapted for liquids and gases. Instead of tracking the acceleration of a solid object, the equations track how a bit of fluid moves, deforms, and flows as different forces push and pull on it. The reason these equations are so notorious, and the subject of a million dollar mathematics prize, is because of turbulence. Smooth, layered flow, or laminar flow, is relatively straightforward to calculate. But once the fluid speeds up and turns into a chaotic turbulence mess, the mathematics become incredibly non linear. The fluid's movement feeds back into itself, creating unpredictable, swirling eddies that become so complex you baisically need a super computer to simulate it. Ok so what the hell is the millennium prize then? The millennium prize is for proof that the equations can result in a singularity or what you see referred to as a "blow up". This is where the equations essentially go to infinity. Ok so why doesnt your cup of tea blow up or turn into a singularity that swallows your house? The Navier Stokes prize isnt about proving that nature makes blow ups or singularities possible, its about proving the limitations of the model's mathematics. Ok so why can a model allow for things we dont observe? This is THE SAME type of mathmatical allowance that general relativity gives when the model breaks down around black holes. The chance that a black hole is an actual singularity where everything goes to infinity is basically zero. Thats not how the universe actually works. Ok so WHY is this possible in the model? The Navier Stokes equations assumes fluid is an infinitely divisible, perfectly smooth substance. In the math, you can zoom in forever, and it remains a continuous fluid. But that's NOT how a real fluid in reality is structured. Real fluids are made of discrete, individual molecules bumping into each other. Theres a real limit to how far you can zoom in. In the proof and the model it means that as a vortex shrinks and speeds up, it eventually gets smaller than the distance between individual water molecules. This cant happen in reality. So the millennium prize isnt about finding if nature allows for infinite turbulence or eddies until a singularity forms, its about FINDING THE LIMITATION OF THE NAVIER STOKES EQUATIONS, proving global regularity false. Ok so why all the drama with the Open AI solution? It used a method called "forcing". In the simulation they essentially used a continuous force to keep the vortex spinning faster and faster until it blows up, the vortex is smaller than the actual molecules that the fluid is made of. It goes to infinity in a time period that ISNT infinity. While the millennium prize doesnt specifically PROHIBIT forcing, most mathematicians would agree that they are specifically interested in the unforced proof, the spontaneous blow up. But that doesnt mean the result isnt legitimate. I hope this clears up some of the misunderstanding.

how has the uk ended up with no frontier lab of its own. we have the talent, the best universities, and london sitting there as the pipeline from research to funded company. the vcs are here. every part of it is here. but we have the highest industrial electricity prices of any developed country, four times what american firms pay and 50% more than france, and we leech off french nuclear to keep the lights on. and where’s our mistral. ai? we had one. deepmind was founded in london in 2010 and sold to google in 2014 for about £400m. we invented the thing and flogged it before anyone realised what it was. this might be britain’s biggest fumble.

1 petabyte of data, precomputed across all ~9 billion possible single-nucleotide variants (SNVs) across the entire 100% of the human genome. for comparision, AFDB is around 23-30 TB and contains ~1.5–2% protein-coding portion of the genome, THIS IS HUGEEEEEEE

For non-engineers the Navier-Stokes equations describe how fluid moves through 3D space It’s the math behind basically all of aerodynamics Engine performance, wing design, fuselage shape, you name it

Some thoughts on the Navier-Stokes news from a professor of aerospace engineering: 1) take a deep breath 2) posts indicating we "solved Navier-Stokes" are incorrect and overblown. A very specific mathematical proof involving a niche scenario for Navier-Stokes has *potentially* been shown (that a perfectly smooth incompressible initial condition with finite energy could produce a singularity--note the assumptions here piled on to an equation set that already involves some assumptions). 3) there is no general closed form solution for N-S and that's not what this news is about 4) this is more of a mathematical curiosity than anything else and could matter a lot to mathematicians but basically changes nothing in the way N-S are used in practice 5) there are already very useful exact solutions to Navier-Stokes that greatly simplify the equation set with the right boundary conditions/geometry like laminar flow through a pipe or a 2D Taylor-Green vortex 6) in aerospace, we already treat N-S as an approximation in many instances and are well aware of its limitations. we don't outright solve these equations anyway and regularly chop off terms or simplify parts (at the expense of accuracy) to make them easier to deal with. 7) to be honest I have never found this specific problem to be particularly interesting because we know a core assumption of N-S is a "continuum" fluid where you ignore molecules and assume hydrodynamic scales >>> molecular scales. We know a real fluid cannot have infinite velocity or energy. But clearly you can push equations outside of their bound of validity and make them produce funky results and singularities.

Absolutely suggest you read the full report, some great details