Alex Mizrahi

@Love2Code DeepMind apparently explicitly trained on pixels-to-svg or some such, would be interesting to compare. Although I'm not sure it transfers to 3D.

I tried to get Claude Code to recreate the Quake start map in my game engine but it really struggled. I even gave it access to a screenshot feature and the original Quake map source. Neither was helpful. I think CC can't process visual information very well.

@monkey_reg @neocentrist the most efficient way to learn. Also some architectures : e.g. there are no LLM relying on small custom specialist modules. Even though it's demonstrated to be possible big labs are just more interested in general e2e approach.

@monkey_reg @neocentrist significantly smarter just by thinking. Everything we've seen so far seems to be in line with that, but it's not really a strong indicator: e.g. there's no reason to believe that MLP is the most efficient unit of compute (might be just the most convenient) or that SGD is

People are in agreement that we're in a short-timeline slow-takeoff world now, right? AI is clearly at the level where we'd expect FOOM to start, but it has not

@vladuhat999 @flowersslop soft prompt. Is that neuralese? Obviously, yes. So what makes it "neuralese", a constrain on information flow?

@vladuhat999 @flowersslop The difference is blurred when you can fork "agents", use soft prompts, manipulate KV cache, use steering vectors, etc. E.g. suppose a subagent gets a KV cache, is that neuralese? "No, that's just a fork". OK suppose we use some hypernetwork to translate KV cache into a

I still dont buy that english cot is better than neuralese cot, except in terms of interpretability which should be subordinate to performance, not the other way around

@Leishman > No one can decrypt it early, not even us. ... > This is a trust-based oracle. Use at your own risk. ??? You have access to the server, right, so you have access to all the keys, no? > a backend KEK that only the worker process uses. Where does backend get it from?

A side project I've been working on is a time-lock encryption oracle that can be easily used by humans and agents. Use it for delayed data access, embargoes, sending messages/files to the future, or anything else you can come up with. 1. Timelock a file in the browser by choosing the unlock time, drag and drop the file, and click encrypt. Easy. You then have the encrypted file to share with others. 2. When a key's time arrives, anyone with the encrypted file can decrypt it in their browser. All of the above can also be done by developers and agents in the terminal using only curl and openssl, which all machines should have installed already. Get your agent to experiment with it! It works by publishing an RSA key for each minute for the next 30 days. The system then releases the corresponding private key at the top of each new minute. It was designed to be maximally simple and compatible with all systems. This is not a commercial project and is not related to @River. I just wanted something like this exist on the internet to see how people use it. Have fun!

@VictorTaelin Hmm, why not IDE? I think it's a good choice for most devs. IDE subscription is a better business model for Cursor than hoping that their model is the best model this week. They can train it specifically for their harness, adapt harness for the model, collect data, etc.

Deleted again because misinformation 🥲 Gemini 3.5 Flash *is* available on the API. Yet, both the API and the CLI versions are 3x slower than on the IDE! See the video below. → Antigravity IDE: 4 seconds (smooth) → Antigravity CLI: 15 seconds (buggy) So the point holds: they want you to use the visual IDE. Problem is: it is 2026. NOBODY should be using IDEs anymore. Get over it. Let it GO. I’m certainly not launching a VSCode fork to use a model, no matter how great it is. They invent a portal gun, only to lock it behind a taxi subscription, because they completely fail to realize their very product deprecates that other thing they think will make them money? Cursor is a great example of a company that (sadly) is very likely fail because of that mindset. Composer is actually surprisingly good model. They should put all efforts in serving it. Yet, they keep locking it under an old school product that nobody wants to use. And even these who DO use IDEs probably won’t necessarily pick YOUR IDE. And they shouldn’t. You do NOT need them to, to make money. Your model is the product. You keep chasing old business models. Completely out of touch. Meanwhile Anthropic is all charging at full speed to sooner or later surpass Google by just serving great models under an API /ctrlv

@VictorTaelin Hmm, they used to offer fine-tuning for Gemini models. Imagine Flash adapter to work with your code base specifically. And I guess for Google it's not a problem to make codebase-to-finetune service...

The new Gemini 3.5 Flash solved the HVM3's wnf bug in 1/3 attempts. This is my main test to take a model seriously. So far only the big models like GPT 5.5 solved it. And seems like it is 20x faster than Opus 4.6 ! Promising but Google will still find a way to fuck up

@oguilhermemora @VictorTaelin Previously Flash was actually more competent at using tools than Pro (as well more mentally stable, etc.). DeepMind is really struggling with post-training Pro.

@VictorTaelin Quando abri o antigravity hj, envie o prompt e vi o agent trabalhando de forma fluída no ambiente, buscando contexto, usando tools e pensei "caralho, tô com o opus selecionado" aí fui ver era o gemini flash kkkkkkkk Não tô nem zoando!

@tszzl It's a scenario which Christiano described in "What failure looks like", but you're talking about it in a positive sense for some reason. This is bad unless you guarantee that ASI follows CEV somehow.

on some level if you want civilization to ascend to a new level you need your AIs to do things that are not legible to you and maybe not even strictly obey you, in the same way that if you hire a great new ceo you give them a lot of autonomy to transform the company according to their own plan, even one which may not immediately read as a winning strategy (imagine the board of directors of Apple firing and rehiring Steve Jobs years later - except the board of directors are chimpanzees) all else equal, companies and organizations that hand more of themselves over to machine intelligence will outcompete ones that demand the corrigibility and legibility tax of human oversight and human design. it is not a stable equilibrium and requires some sort of vast cooperation scheme if you’d like to enforce it real asi alignment has to operate at a deeper level than oversight, control, or human corrigibility

@nabeelqu @GrantaMag poetic text. I wanted something more concrete. But if you want a metaphor-heavy text it's probably appropriate.

@nabeelqu @GrantaMag This sort of style of writing was in use for thousands of years. Of course it's a trope. But should we all speak like gen-z now to avoid repeating old style? I remember "Not X, but Y" construct irked me when I read 12th century Ukrainian epic as a kid. Very metaphor-heavy

Well, this is a first: a ChatGPT-generated story won a prestigious literary prize (The Commonwealth Prize). "Not X, not Y, but Z" sentences everywhere, the "hums" trope, and plenty of other obvious markers of AI writing. A major milestone for AI, at any rate... @GrantaMag

@ercwl @tegmark Qualia also rather obvious: sensory information which is connected to the world model is different than some abstract data.

@ercwl @tegmark I think it's just common sense that a creature which needs to execute complex, dynamic plans would need to model the world and itself in the world. And sensing yourself in the world is what consciousness is, by definition.

most interesting thing I’ve thought about this week: @tegmark’s theory that consciousness is the most efficient way to implement higher forms of intelligence slightly diff take than consciousness being an emergent property (which makes it sound accidental rather than necessary)

@IsaacKing314 Looks much less dangerous from fairly typical driving scenarios (e.g. incoming traffic on a narrow road with no separation).

Ok I'm sorry but this picture is exposing the wordcels. There is ~zero risk that someone in this position could accidentally fall in. She's 3 feet away with a low center of mass and the skateboard wheels aren't pointing towards the edge. That is just not how physics works.

@0xdoug can't get rid of a clanker accent so Eloi can pretend they can do something uniquely beautiful. Then everyone will be happy

@0xdoug There might be even some conspiracy going on: big AI labs don't want to dispell "AI is bad at writing" myth because a lot of people are dreaming about writing in a world of abundance. So they want to leave writing as an inherent human activity. Clankers will pretend they can't

Even 30B models are crushing grad level math. The hard to escape conclusion is math isn’t actually that hard. Humans are just really bad at it. Writing a 40 page short story with narrative consistency probably requires more intelligence than winning an IMO gold medal