people really have lost their basic sense of manners while speaking to elders, especially teachers. i agree some teachers might not be good, doesn't give me the right to be rude to them, they're still teaching people or have taught people, and people should respect that

I’ve put a ton of tokens through Jev now, and...I have thoughts. 1. The big labs lost the script TypeSafe’s intro video says, “We’re building prod, not God” — and can I just say? Hell yeah. @typesafeai built something for engineers like me, so I can build products for people like your mom (she says hi btw). They’re not trying to create “Machines of Loving Grace”; they’re trying to build things that let OTHER HUMAN BEINGS create new types of products. The proof is in the pudding. When ChatGPT shipped, we were all blown away by…ChatGPT. When Jev shipped, we were all blown away by what everyone was making with it. I cannot tell you how refreshing this is. It takes a sincere form of humility to believe that you, the creator of a technology, will not be the best at deploying the technology into the marketplace. The big labs, especially Anthropic, have proven not to have this humility…and now everyone hates AI. Thanks, Dario. I guess what I’m saying is: Diogo for president. 2. Ultra-smart classifiers were genuinely a missing primitive. I don’t know how this got missed. Diogo calls it “machine-native intelligence,” and that’s really what it is. I cannot tell you how much bending and twisting I've had to do with LLMs to get them to act like a classifier when they just weren’t. I’m absolutely shocked this is the first time this “type” of model has come to market. These models will unlock AI utility for entire industries: robotics, trucking, aviation, meteorology, finance, defense, retail…even your smart coffee mug is going to use this. 3. It’s still too expensive Jev is really, really cheap compared to LLMs. But you don’t use it like an LLM. In my full self-driving example, or any robotics example, you’ll be making decisions several times per second continuously. The #1 value proposition of a model like this is speed and cost. To be clear, at current pricing, the model will still be successful and well integrated. It will be able to replace a number of tasks that were already being performed (poorly) by LLMs. However, to unlock industrial-scale demand (Jevons paradox), I believe it needs to be roughly 10x cheaper. Currently, Jev is $0.042/M input, which is actually 7x more expensive than a cache read on DeepSeek V4.1 Flash ($0.006). If our industrial application requires decisions at 10 Hz and we provide only 10k of context, that would cost $0.0042/second to operate, or $15/hour, $600/week, etc. The good news is Diogo said on a recent AMA that they have substantial margin at their current pricing and have already considered dropping the cost further. I hope it's a lot further. 4. Latency is a real limiter I really hope TypeSafe AI works with an edge network provider to improve the co-location of models and reduce network latency. Ultimately, these models really need to run on-device. I would even be willing to pay some kind of recurring license fee to have a great closed model running on my own hardware, just so I can reduce the TTD (Time To Decision) as much as possible. Imagine running a model like this at 60 Hz, or even 120 Hz. At that point, we’re looking at a new kind of logic gate — we can’t even imagine what that will be like. In summary, this one really is a game changer for everyone building AI products, and, assuming cost and latency are further improved…it’ll be a game changer for everyone doing anything.

Well, a week late - but better late than never. Had a blast of a podcast with @sunnyray - my first one as well ^_^ . And, as I always say - The Model was never the bottleneck. Link in replies :)

Why would I deceive you @elonmusk ? #scamaltman

meet @mostik_ai! what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can't say much about while the competition is still running. and this, which I can. everyone's arguing about whether open models will catch up to frontier models. we think it's the wrong question. here's the one we pose: why does a frontier model have to generate your answer at all, when the only thing you need from it is the reasoning? we do this by enabling models to communicate in latent space. through our protocol, hidden states pass straight from a frontier model into a small one running on your infrastructure -- no text between them, and neither model is fine-tuned. two models from different families, sharing reasoning, both left untouched. how do we know it works? we tested it on a setup where a 753B model reads the problem, and a 4B edge-class model writes the answer. with this approach, we get results 80% as accurate as the frontier model, but at 20x faster performance. we're committed to preventing frontier model lock-in and are already partnering with inference providers to accelerate open-weight adoption. we've done this between 15 of us, in four months, 12 PhDs and a Fields medalist, backed by @generalcatalyst WIRED has the first external account of the company and the work: wired.com/story/russian-star… full writeup, the setup, and all the numbers: mostik.ai/read-more

GOOGLE HAS BEEN FAILING PRETTY BADLY TO PUT OUT A MODEL THAT CAN ACTUALLY HANG WITH THE BEST LATELY tested Gemini 3.8 Flash against Kimi K3 on the same Sticky Ball game Flash was insanely fast and used way fewer tokens, but the actual game was nowhere close. K3 had much better mechanics, movement, and overall game design Flash clearly has the speed and efficiency part down, but the gap in what it can actually build is pretty big

so when did you start coding? answer in BC = Before ChatGpt/Claude me 0 BC

We just evaluated Gemini-3.8-Flash on our @PhyseraAI - TB bench. Our analysis on 5 random tasks from the set: 1. It is excellent at deriving and implementing a coherent numerical method. It turned an internally consistent mathematical model into working code, especially when it can derive a checkable optimum. 2. It is inconsistent on edge-case semantics. It knows the individual primitives, but assemble them in the wrong order when a specification has interacting edge cases. 3. It tends to overbuild static-analysis solutions while missing the hardest coverage cases. 4. In multiple tasks long trajectories do not imply better outcomes. More exploration became speculative scope expansion rather than targeted verification. I think it is promising low-cost choice for numerical/scientific coding / transformations with crisp formulas / tasks where it can independently check residuals or invariants but had fallbacks for production shell tooling / static analysis / clinical derivations / compliance-style work.

Deepseek V4 Flash 0731 is dollar-for-dollar ~3x better than GLM 5.3 Flash. - It shows ~33% better alignment towards changing objectives - It produces ~17% more output tokens over GLM 5.3 Flash - GLM Flash performs ~37% better for STRICTLY coding tasks. Tested with @opencode.