Bala

Lifting weights 💪, biases & activations @nvidia. @CarnegieMellon Alum.
Can shuffling some bytes summon a silicon god?
Opinions are my own.

[Bala @bala_bmm]

So unfortunate that Helium isn't the real laughing gas

HeHeHeHe

[Pushkar Ranade @magicsilicon]

Doping silicon is the only time ‘adding impurities’ is considered dope.

[Bala @bala_bmm]

OpenAI (co founders) will be probably pump Cerebras and make a strong exit from their Cerebras seed investment by the time Jalapeño comes into production use

[Patrick C Toulme @PatrickToulme]

Cerebras is going to 1T. Deep in the GPT 5.6 announcement is this gem. This offering will be a high priced fast mode offering for agentic coding. You will see customers like quant funds etc who have unlimited budget and the need for the fastest agentic coding in the world using GPT 5.6 on Cerebras. GPT 5.6 on Cerebras will enable coding projects to be completed in minutes due to this decode rate. The rate of progress in technology in the world is about to accelerate. $CBRS

[Bala @bala_bmm]

deep

[Charles 🎉 Frye @ AIEng World's Fair @charles_irl]

I didn't have time to write you a short PR, so I wrote you a long one.

[Bala @bala_bmm]

Holy smokes!

[Noam Shazeer @NoamShazeer]

I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there.

It was a difficult decision to move on. I’m incredibly proud of the amazing team at Google and everything we’ve built together. It has been an honor and a pleasure to work with all of you.

[Bala @bala_bmm]

time to digest the fact that nobody is entitled to frontier tokens in frontier domains. and it won't matter in the long run anyway. it's a fun time till then

[SemiAnalysis @SemiAnalysis_]

BREAKING NEWS: Anthropic's latest model will NOT help you if it thinks your ML research/ML engineering is interesting, and/or will secretly degrade its IQ so that the average engineer won't notice. We are already seeing Anthropic's latest model's moderation filters our GPU inference research and programming 😭

[Bala @bala_bmm]

when you know Andrej is probably going to A\ as Hamza but can't prove it

[Andrej Karpathy @karpathy]

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

[Bala @bala_bmm]

like the last few years, just before the new Gemini launch

[leo 🐾 @synthwavedd]

🚨 SCOOP: The development cycle for GPT-5.6 is now in full swing at OpenAI.

The first checkpoints of the model began testing internally over the last few days, with a release likely next month.

[Bala @bala_bmm]

yes. one step further - specialized models for user facing and worker agents.

user facing agent just needs to be fast, orchestrate well and reasonably grounded, update html.

worker/background model - need not be fast, but should be grounded and intelligent maxx that's all

[Thariq @trq212]

HTML is the new markdown.

I've stopped writing markdown files for almost everything and switched to using Claude Code to generate HTML for me. This is why.

[Bala @bala_bmm]

gradient descent over everything

[Lisan al Gaib @scaling01]

I think returns to intelligence are nonlinear because decisions are path-dependent

early choices in code, experiments, or strategy can compound positively or negatively over time

for example by avoiding dead ends or preserving optionality

it's why I am a big fan of very long running tasks and massive benchmarking budgets

GPT-5.5 and Mythos Preview are only marginally more intelligent than previous models and have pretty much the same performance up to 10M tokens, but after that they go absolutely ballistic

[arb8020 @arb8020]

gpt-5.5 prompt for codex seems to have a duplicated line trying to get it to not talk about creatures?

Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query.
[...]
Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query

gh link:
https://t.co/1LF8FkRaVf

[Bala @bala_bmm]

@benitoz @PatrickToulme imo abstractions are there for a reason, they provide guarantees and set expectations. I see different abstractions as different tools available for agents to use. Unless absolutely bottlenecked by higher abstractions, there my not be a need to circumvent it for everything

[Bala @bala_bmm]

nobody's handwriting kernels, so this makes sense.... adding an agent friendly skill would be 💯

[Patrick C Toulme @PatrickToulme]

Launching pyptx — a Python DSL for writing NVIDIA PTX kernels.

One PTX instruction = one Python call. Write pure PTX in Python.

Direct Hopper + Blackwell support: wgmma, TMA, tcgen05, mbarriers. JAX + PyTorch integration.

Includes GEMM, grouped GEMM, RMSNorm, SwiGLU, and a PTX→Python transpiler

pip install pyptx[torch]
pip install pyptx[jax]
https://t.co/PcISpsaeQ5

[Bala @bala_bmm]

custom app dev in 2026:
logic = skill packaged with harness
UI = tui

[OpenRouter @OpenRouter]

Introducing "create-agent-tui"

A skill for building your own agent harness + terminal UI (TUI). The skill walks you through 4 different ways of customizing the look, and supports dozens of optional features 👇

[Bala @bala_bmm]

@steipete @WG_Mojo @openclaw +1 workaround is to set a higher effort level in Codex. that makes it keep going on without stopping

[Bala @bala_bmm]

@fcoury native markdown preview when?

[Bala @bala_bmm]

codex acting lazy seems like a direct effect of having reasoning levels medium and high, extra high.

user <-> model reasoning should always be high, only multi-step inference/reasoning should be controlled by the levels.

understanding user's intent >>>> everything else
@gdb

[Bala @bala_bmm]

forget benchmarks, a claude as smart as Codex/GPT 5.4 is all that is needed. Is Mythos up there?

[Alex Albert @alexalbert__]

We released Claude Opus 4.6 just two months ago. Today we're sharing some info on our new model, Claude Mythos Preview.

[𝚟𝚒𝚎 ⟢ @viemccoy]

The weather in San Francisco is beautiful right now. It is the perfect day to use tmux at the park on your MacBook Neo to oversee your agent harness.

[will depue @willdepue]

you guys really went easy on Project Hail Mary. just arrival for retards. interstellar for chuds