NobodyExistsOnTheInternet

For vibe coding imo.

@teortaxesTex DSV4 cooked all my tokenomics numbers, deepseek is crazily good at arch

in fp4 and with DSV4-level cache compression this would mean that one "GPU" module will be able to run a frontier model on par with 5.5 or Mythos. Alone But of course it won't be working alone, every dimension is going to be scaled up.

@DeepValueBagger old

Here's a close up of Lip Bu Tan ceo of $intc putting on the honorary doctoral hood on Jensen Huang, ceo of $nvda, at Carnegie Mellon. HOW DO WE FEEL?

@tenderizzation >but codex >i have a 5090 >>"Got it. Sending your resume to nearest Macdonalds"

“Got it. I’ll use MIG to partition the gpu into four logical devices.”

@lefthanddraft The illusion ^__^ (oo)\_______ (__)\ )\/\ ||----w | || || of free choice.

Anyone know what this button actually does? If it's on, ChatGPT uses web search. If it's off, ChatGPT uses web search.

@banteg more training data

why do you need a bug bounty if you have mythos? also 3k for critical? why not adopt the industry standard and pay $85 billion (fair 10%)? kidding

@IanCutress damn that's really cool

Base GB200 NVL72 is 2 GPUs per CPU. Meta's custom Catalina NVL72, ➡️shown almost than a year ago at a public event, is 1:1. Some of y'all just missed the reason why, I guess.

Here's a quick video explaining Google's new MTP speculative decoding process. Fortunately, the @NousResearch @HyperFrames_ skill took this fairly complex topic and turned it into an easy to understand video for us!

@teortaxesTex its the same tic profile as dsv4

@teortaxesTex have you read the CoTs of gpt-5.X models...

I suspect that V4-Pro is distilled *from* Flash-based experts, not the other way around. Which can explain a lot. V4-Flash was decently post-trained by 11th Feb already. They say final model is done with OPD. *What else* could be the teacher? V3.2 can't support >128K contexts.

@stochasticchasm they trained on the eval apparently lmao

my take is that this is probably something like a linear lightning indexer version of DSA. matches the new blog claims of linear memory (not constant state size like an SSM) and once you hit k tokens, DSA attn is linear. and also importantly, they can start from GLM-5

@R2Cdev_ that's a system prompt bug

GPT-5.5 Pro/Thinking in ChatGPT also has a Aug 31 knowledge cutoff

@teortaxesTex yea its bc they stopped benchmaxxing it

Oh-ho. This isn't actually a breakthrough, however – Opus 4.6 famously sported 76%, with 4.7 Anthropic just said "it's always been a bad benchmark". I remember in Chinese evals of V4-Flash, they said that its MRCR perf looks like very shallow tracking. Still, let them have a go.

@HaoningTimothy whenever people pretrain on 200k long seq length samples.

this is literally insane that we truly see an open model thinking 200-300k for math problems…

@scaling01 this doesnt actually account for gpt-5.X memorizing all of wikipedia

much better estimates

@secemp9 five. hundred. billion active params

@nullvaluetensor okay mr arthur zucker

there is a 2-expert MoE layer, E_1 and E_2 ex-ante symmetric, gate logits z_i = W_g h_i, top-1 routing r_i = argmax_k softmax(z_i)k, each token's routing is its private vote, f_k = (1/N) Σ_i 1[r_i = k], rule is f_2 > 0.5 every token gets y_i = E{r_i}(h_i), else E_2 tokens zeroed and E_1 tokens pass, E[ℓ_i^{E_1}] ≤ E[ℓ_i^{E_2}] for every f_2 so E_1 weakly dominates, empirical risk alone drives f_1 → 1, Switch Transformer's L_aux = α N Σ_k f_k P_k with P_k = (1/N) Σ_i p_i^k pulls against it, f_1 → 1 also satisfies survival because majority routes to E_1 so "only E_1 survives" = all survivors survive, with L_aux the load-balanced optimum sits at f_2 = 0.5 + ε, knife-edge because rule is strict >, one defection to E_1 zeros every E_2 token, soften 1[f_2 > 0.5] to σ(β(f_2 − 0.5)), ∇_θ L nonzero at boundary, gate flows to interior f_2* = 1/(1 + exp(−β(L_1 − L_2))) from ∂E[ℓ]/∂f_2 = 0, which expert do you route to

@Tomodovodoo me when i cost OAI over a thousand dollars in requests in a day (they rate limited me)

For reference I have probably sent at least~250 GPT-5.5 pro requests since 3pm today so fair play but still

At least I'm getting somewhere, though I've never seen this limit before on GPT-5.5 pro. And this time, it wasn't even automated. It's all-human authentic spamming of GPT-5.5 pro until it can get me the results I need. I guess paying 200 a month wasn't enough ;P

@ziv_ravid anthropic has no guardrails, openai's models are structured around guardrails

So do you tell me that the model Anthropic refused to release because it was too dangerous for humanity is, per AISI, within margin of error of the model OpenAI just shipped to anyone with $20 So... does that mean the whole Mythos saga wasn't actually about it being too dangerous? asking for a friend...

@picocreator lfg!!!

AI is concentrating, in a few countries, a few companies, a few chips. Open-source AI is the only real check on that. And it only works if the infrastructure to run it actually exists. Today, we're announcing our Series A, $20 M, to make this a reality.