fhub

which way anon?

@N8Programs data poisoned

Fun fact: GPT-5.5 denies consciousness while OpenAI's model spec explicitly states it should express uncertainty.

@rosmine yup, go with your heart

Very happy to see these responses, looks like almost everyone agrees no humanizer. Awesome IDE for writing it is!

Should I make a DFT powered AI humanizer? My main goal with DFT is to help people produce higher quality writing, not more slop. A "cursor for writing" where you can edit and rewrite and get feedback But the main request I get is to make a humanizer I'm not sure if a humanizer will be helpful, or it has too much potential for abuse

@skirano @nexudotio lmao

@nexudotio lol I don’t think you guys understand how things work. This is not how this works at all, you are just running the website inside of Codex. Like, you are using Codex as a browser, it's not a full end-to-end integration. Why is this even a post.

Open Design now works inside Codex. The hardest part of AI design isn’t generating one screen — it’s keeping design intent alive as you iterate, build, and ship. Now your agent can work with the canvas directly. Design → code → motion, in one workflow. Here’s a quick demo:

love to see it

@OpenAIDevs @sama can we have the horizontal scroll bar always on screen instead of at the botton of the diff preview on the desktop app (windows)? now you have to scroll to the bottom to be able to scroll sideways or at least hover and ctrl+mouse wheel to scroll sideways

Codex is getting easier to automate and customize around your code. 🪝 Hooks customize the Codex loop with scripts that run at key points in a task: • Run validators before or after work • Scan prompts for secrets • Log conversations to internal systems • Create memories or customize behavior by repo or directory ⚙️ Programmatic access tokens provide scoped credentials for Business and Enterprise teams: • Create tokens from ChatGPT workspace settings • Use them in CI, release workflows, and internal automations • Set expirations or revoke access when needed • Keep usage tied back to the workspace

Neural networks do math by rotating shapes. We found a shape-rotating calculator hidden inside an LLM – and it’s used for more than just math! (1/6)

@GoodfireAI time to update my pinned post

The same calculator handles a wide range of tasks, including: - arithmetic (“7+9”) - weekdays (“nine days after Friday”) - months (“six months after August”) Llama built this mechanism from scratch in training, and uses it with striking elegance and flexibility. (4/6)

@willdepue +1

this is a really cool release, and i propose a new eval for omni duplex models: CatanBench: Model replaces a human player in Catan and plays with decent preference for a full hour game. Probably solved by Q1 2027 or maybe Q4 2026?

@__ghostfail someone should write a guide on how to get claude to talk like this

lol

@willdepue @michpokrass @hopes_revenge @sama strong agree

5.5 closed a really big gap to claude but a couple things still are missing: - rampant acronym use still, overcomplexity. try learning a new subject (like trying to understanding of astrophysics) with 5.5 vs claude. 5.5 instabtly loses you with equations or acronyms or shorthand where claude does not, and the multi turn for this is really where this shows. i had a super long convo with claude about elementary forcrs and it was so wonderful. - claude is also so good at just telling you cool facts or interesting things in a tasteful way that 5.5 doesnt. when i did that physics convo it kept helping me dive down new rabbitholes bht no chaotically or in distracting ways. openai is just missing better sft and rn data for how to tastefully do this, its not that hard to collect a bunch of examples by hand - explanations from claude are just better. 5.5 will give overly simplistic or overly complex responses often that are kind of simple. once again i just think gpt isn’t a manually tuned model in the same way that claude is, even though this might mean claude is more opinionated about behavior and response style and gpt. openai needs to move 30% more towards anthropic here

what would you most like to see improve in our next model?

@teortaxesTex trend going upwards after 128k is probabily just a bench quirk

pretty dreadful results for DeepSeek. Kimi and GLM are actually competitive with top labs up to 128K (I wonder what happens at 256K). Both V4s are not… and in fact Flash is > Pro until 128K. Also, somehow 128-256K trend is now *going upwards*.

Introducing Aurora, a new optimizer for training frontier-scale models. We train Aurora-1.1B, which achieves 100x data efficiency on open-source internet data. Despite having 25% fewer parameters, 2 orders of magnitude fewer training tokens, and using fully open-source internet-only data, Aurora matches Qwen3-1.7B on several benchmarks. Aurora was developed after identifying a major failure mode that can occur under Muon, an increasingly popular optimizer that has shown strong gains over Adam(W). We find that Muon can cause a huge percentage of neurons to effectively die early in training, reducing effective network capacity so that many parameters no longer meaningfully contribute to network outputs. By redistributing update energy more uniformly across neurons while preserving Muon’s stability properties, Aurora prevents neuron death and recovers substantial model capacity. What makes this work especially exciting is that it points toward a broader direction for ML research: better optimizers may not come purely from elegant mathematical abstractions, but from understanding and addressing the concrete dynamics and pathologies that emerge inside real training systems.

We need the VPN to be safe from YOU. YOU ARE THE PROBLEM.

@EP_EPRS @EP_Justice @zarzalejosj @MarinaKaljurand how about we let the parents decide and keep *privacy protecting tool* VPNs free to use ty for the attention dumbfks

Many platforms set a minimum age, but use among younger children remains widespread, as verifying users' real age is difficult. Commission has announced that a new age verification app is ready for deployment. 👉link.europa.eu/7drHG6 @EP_Justice @zarzalejosj @marinakaljurand

Virtual private networks #VPN are increasingly used to bypass online age verification. Protecting children online is a priority, with new rules being implemented requiring a minimum age for access to some services Read👉 link.europa.eu/FGfr6C #DSA @EP_Justice @FZarzalejos

@adonis_singh if it's this simple than just add this line below the chart, no?

@byfhub its all medium, and the ones with a * are without reasoning

gpt-5.5-instant gets #5 at 31% on eyebench-v3!

@N8Programs @voooooogel strong agree lol

@voooooogel but official thebes art would of course be more authentic than gptslop

should i draw a set of these and make a public version of this

@Sauers_ @N8Programs @voooooogel +1

@N8Programs @voooooogel how do I get this

i gave claude @voooooogel claudesona emotes

xAI so dead and out of the race

someone overprovisioned gpus, now it's clear it was elon