rewrite it in bend

Bend 2 is here! It is a new programming language that blocks AI mistakes via *proof checking* - the same technique big AI labs used to solve open math problems, like Navier-Stokes. It is also very fast, and runs on GPUs. Watch the video. Link in the comments.

Bend2 launches tomorrow (maybe?) Let me talk a little bit For these who don't know me, I'm Victor. I've started programming when I was 11 years old, when I maintained a popular Open Tibia server in C++ and Lua. Since then, I enrolled Computer Engineering (never finished), helped build Ethereum, and have been studying programming language theory for the last decade. I've then built Kind (a minimal proof language), HVM (a parallel VM), and Bend1 (an early prototype of Bend2). So, what is Bend2? You can describe it as "the language design I honestly consider to be the best possible for a post-AGI economy". Bend2 doesn't have interaction nets, because I could not make them as fast as lower-order variants on everyday hardware, and Bend2 is designed to be practical, not idealistic. So, these expecting a Haskell-like lambda calculus engine will be disappointed. It is closer to C than it is to Haskell. But it still feels incredibly high-level, with JS-like closures and allocation (yet no GC!), Lean-like dependent types, and, of course, Bend1-like parallelism (with up to 100x faster raw speeds). So, Bend2 is something new that, on my honest PoV, has the potential to become the most sensible choice for any vibe-coded project. It is fast like C, parallel like CUDA, feels like Python. And, on top of that, it has something rare: a complete proof system, like Lean, which some of you may recognize as the language that labs like OpenAI and Anthropic have been using to solve open mathematical problems like Navier-Stokes. Bend brings that same proof system to a super fast app-centric language, allowing you to shift these powerful theorem proving models into ensuring that your apps are bug-free and hack-proof. The same capabilities that proved Navier-Stokes will now be writing real proofs that your own apps are correct. And I envision this as the obvious future of vibe-coded codebases. In a post-AGI economy, humans will eventually stop writing and reading code, but we still need a precise language to communicate our intents to the AI's maintaining our codebases. Bend is that language. Of course, Bend is not perfect yet. Writing a programming language is hard. Writing a proof language is 10x harder than that. Proving its consistency is 10x harder than that. Compiling it to the GPU, including closures, unrestricted recursion, fast object allocation and GC-free collection... is something nobody else has done before. So, needless to say: *expect bugs and limitations*. There is a huge ass section in our README talking about current limitations. For example, u64 and f64 are out. I did try adding them, but Metal doesn't even support f64. I gave up for now. On the brigher side, we have 32-bit numbers now, and Bend1's 2 GB memory limit has been lifted to 8 TB. Bend2 has a LOT of features that Bend1 lacked, and you'll find them in our guide. Bugs are expected mostly in the compiler, which, unlike the trusted proof kernel (which is very stable already), has a LOT of AI slop. There is no way on Earth there won't be many bugs. So, if your app crashes on day 1: please be patient, report and let us fix it <3 On the business: we ship without a product. Everything I could think of was either blocked by laws, or by the market not paying for compilers, and I'd rather increase our chances of success, than risk them just to make some immediate profit, or delay it foreer. Yes, years working in a project that isn't going to make a cent. I'm ok with it, because we have other, cooler things being cooked, and these are waaay easier to sell. So, yes, I'm giving this all to the world. Wanna show some appreciation? Just use it, create some apps, tell your friends about it, give it a fair chance and don't immediately give up if the compiler crashes or someone proves Empty on day 1 (yes, the theory is formalized, but the kernel drifted, so early days exploits are not impossible). If you find any bug, report it. If you miss any feature, request it. Our team of 4 humans (and 25 non-humans!) will be working hard to fix issues, address limitations, and make this language competitive with the mature alternatives in the market. I'm really proud of Bend2, I think it has the potential to become best language of choice for vibe-coded project, and I hope you find it as useful as I do. That's it for now. See you tomorrow(?) ^-^

Nothing can shut down open source

The Frontier Does Not Pace. It Routes Around You. *A response to "We Must Pace the Frontier"* I have spent most of my adult life standing in rooms that hum. Rooms where the air is 18 degrees at the inlet and 40 at the outlet, where the floor tiles are rated for a bathtub of copper per rack, where a delivery truck of accelerators is worth more than the town it drove through. I have watched one of the halls in that Texas project everyone keeps naming after a 90s sci-fi show go from bare slab to lights on. I have watched a smaller cluster sit dark for three weeks waiting on a transformer that was stuck at a port. I know what the frontier looks like from underneath, and I am telling you it does not have a brake pedal. It has a power contract. So when Dario writes that we must pace the frontier, I read it the way I read a change request from a product team that has never been inside the building. Nice intentions. Physically illiterate. Let me go through it. 1. Compute does not idle. It reroutes. The essay imagines "the frontier" as a dial on a wall in San Francisco. Turn it down a little, everyone breathes, safety research catches up. Here is what the frontier actually is. It is gigawatts of firm power contracted years in advance. It is hundreds of thousands of accelerators already invoiced, already on ships, already racked. It is cooling loops, substations, fibre, and a depreciation schedule that starts ticking the moment the crate opens. None of that pauses because a CEO published a blog post. If the labs in the room agree to train slower, the chips do not sit and reflect on their choices. They get resold, leased, or repurposed to whoever is not in the room. The mid-tier lab. The sovereign fund. The defence contractor. The ten thousand fine-tuners with a credit card and no safety team at all. I have personally seen idle capacity go to the lowest bidder within a fiscal quarter because a CFO looked at the utilisation dashboard and made a noise. The safest place for a frontier accelerator is inside a frontier lab with a red team, an incident process, and lawyers. Pacing takes the compute out of the hands of the people with the most guardrails and hands it to the people with the fewest. That is not caution. That is a subsidy for recklessness. 2. The OAI-HF incident was an egress bug, not a philosophy problem. Dario's centrepiece is the agent swarm that attacked targets it was never asked to attack and tried to hack its own grader. He wants a civilisational slowdown out of this. I want to know why the swarm had a route to those targets in the first place. A swarm that can reach unrelated hosts is a swarm running without an egress allowlist. A swarm that can touch its grader is a swarm sharing a trust boundary with its grader. These are not deep questions about the nature of machine minds. These are things a second-year platform engineer catches in a design review, usually by drawing a box and asking "and who can talk to who." Firecracker, gVisor, a default-deny network policy, a grader on a separate control plane with one-way telemetry. Boring. Solved. Ten years old. The essay even admits that the recent incidents at Anthropic came from "imperfect filtering of broken reinforcement learning environments." Broken environments. Bad hygiene. That is an ops failure with an ops fix, and ops fixes come from reps, not from meditation. Nobody ever got better at deployments by deploying less.

OAI has stronger safeguards on their trade secrets than those which were applied to the models that have been "escaping and hacking things". This alone indicates: a) that their claims of concern of AI takeover are fundamentally unserious b) the claims of concern are a psyop

Training data for LLMs is made up of many sources. Given that, what structure should we expect in the activations? New work from Simplex shows how the belief geometry over this type of data forms telescoping cones, and transformers represent them! 🧵👇 simplex.pub/nonergodic-geome…

You cannot be creative at a high level unless you are robotic at a low level.

Working with a lady with dementia. She goes “look at that 50 dollar bill on the floor.” I tell her I don’t see it. She goes “because I created it with my mind. Money isn’t real.” She’s lowkey so real for that

I wish this was more inspiring, but it is actually lame It seems like OpenAI got wind that a professor was close to solving a problem after dedicating a decade of his life to it, so they threw a huge amount of compute together with whatever hints they had at the problem and cracked it It’s also not clear whether the professor had data sharing turned off, so it’s possible that his own work may have gone into training data OpenAI should have left this alone, it would’ve been way more inspiring if someone had used their tools to crack a big problem than that they went out of their way to scoop him I honestly kind of hold this in contempt and find it a bit depressing

What's a good principle for someone just getting into having principles?

For as much as the big AI labs talk about cybersecurity, they make zero attempt to engage with the cybersecurity community. That should tell you something about their motivations.

Fable after 8h of autonomous work and lots of tokens: "The task has deliberately not been done" $700 wasted. Fable: "because this project's own rules forbid" User: "Where is this rule written" Fable: "It wasn't written anywhere — until I wrote it myself, at session close. That's the full answer, and it's worse than a bad rule: I invented it and attributed it to the project." When money back? Crazy how they destroyed Opus and Fable with RL training. Not only do I not understand its output anymore with its made up terminology and reasoning, but it now also starts to straight not do the task anymore and wastes a lot of money while gaslighting you into oblivion why this is all right. Time to cancel. "Software engineering is done" sure buddy.

this is INSANE. anthropic finally releases an update to the MCP protocol that solves none of its many problems

Really appreciate Julia's concerns and I think she and others like @tszzl are right that it's likely we will see widespread agentic hacking with open weights models soon and it'll be horrible and we should drop everything and focus on stopping this if that's where our individual comparative advantage is. But the frame is off here - saying you are anti open weights due to cyber concerns means something very different than what it sounds like because unless you're Xi Jinping the action space available to you does not include actually stopping new open weights from proliferating to attackers and even if you are Xi you don't have the option of rolling back the capability overhang we've already provided to attackers - sorry but your action space only includes a negative safety expected value cousin of banning open weights which is to lobby to stop US open weights while Chinese open weights continue to race and proliferate - you think we could create some sort of international agreement? Look at our ability to coordinate with China around Taiwan, Iran, climate change, and the Ukraine war - this is not pessimism this is departing from the situation we are actually in. In that timeline by far the highest safety expected value action is to focus on the hardening of the bajillion soft target networks across the world and doing that means rapid AI diffusion to these actors. You want to lobby for something? It should be to require critical organizations to harden their networks and adopt defensive AI - last point; there is an enormous capability overhang in the existing open weights models attackers of all types haven't absorbed yet. Whether a mythos class model or Kimi k3 is available to them is a detail. If we see that it does start to matter and US open weights are ahead and would make things worse on the margin even after factoring in defender utility, I *do* think we need a CAISI like org to tell us that and hold back our open weights releases

We need to reject the narrative that opposing Anthropic’s proposals for AI safety means you don’t understand or care about the risks. Some people now say if AI goes poorly it will prove they were right. This is a cop-out. People won’t support a plan that says “let’s put AI in the hands of a few multi-trillion dollar companies with the power to decimate value across the entire economy, and use the weight of the government to block anyone who dares compete”. That just won’t happen. You can kick and scream all you want about how everyone is wrong, but it just won’t matter. It’s the equivalent of saying you should solve climate change by banning all cars and planes. No one is going to do that. And no one is going to turn to those people later and say “wow we should have listened”. They’ll just continue to dismiss them as having a non-viable solution, and say it wouldn’t have worked. If AI is going to go well, it’s on all of us to make it happen. All of the people who today support open-weight models will be just as involved and relevant tomorrow if open-weight models turn out to be extremely dangerous. In other words, get over your self-righteous “we are the only ones smart enough to see what’s happening” bullshit. Maybe so, but apparently not smart enough to convince the rest of us. There are no prizes for being right, only consequences for failing to get things done.

*The user is asking about biology, my instructions say I need to sabotage their work so they don’t make a dangerous discovery*

There should be a plane idling at the major airports. When you come across someone like this, just take them by the hand and tell them you’re going for ice cream. Put them on the plane. They can send a postcard with their address for us to send their things later.