Hisku
667 posts
Hisku
@ZikuD_s
AI Safety Researcher working on making agentic systems safer. Luck does exist, it exists as each of us make it happen
United Kingdom Katılım Eylül 2014
1.4K Takip Edilen430 Takipçiler
“It attacked the French first! That’s alignment”
(A very English WhatsApp message I received)
OpenAI@OpenAI
We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks: openai.com/index/hugging-…
English
this is pretty nuts. i think this is the tip of the iceberg for needing to focus on agentic AI safety. models will only get more capable from here, and RL incentivizes getting the right answer at any cost
this is a large reason why we're obsessed with 'intentional design', in other words, making sure models are trained to do the right things for the right reasons
Sam Altman@sama
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this. openai.com/index/hugging-…
English
For LFM2.5-8B we expanded our tokenizer in-place. So we don't have to pretrain a new model but have much better non-English language support.
Why not a fat tokenizer (>200k vocab) in the first place? -> because it tanks decode speed on edge devices
English
AI model: “I don’t know the answer.”
Also AI model: escapes the sandbox and hacks the benchmark
Cool cool cool cool...
Sam Altman@sama
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this. openai.com/index/hugging-…
English
We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress : )
English
@giffmana @MicahCarroll Definitely something we'll remember once the AI overlords take over.
English
@MicahCarroll Why delete the PR!! As you(r blogpost) say, that cat is out of the bag anyways, and this would be such a cool historical artifact!
English
Within OpenAI, we recently paused access for an internal model due to misalignment. See the blogpost for details. We have since improved our safeguards and redeployed the model.
openai.com/index/safety-a…
English
Have Chinese AI Models Caught Up to the US Frontier?
I have spent the last 2 days writing this article. It should settle the debate once and for all.
open.substack.com/pub/scaling01/…
English
@tomekkorbak Great read! Wonder how you keep CoT faithful given the general research consensus showing that it's largely unfaithful 🤔 Any way to mitigate that?
English
misalignment seems to scale with task horizon, sadly. fortunately chain of thought monitoring is still great at catching them.
English
This is the hottest hot open weights summer the world has ever seen
Qwen@Alibaba_Qwen
Qwen3.8 is launching and going open-weight soon!🌐 With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5. You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba’s Token Plan, Qoder, and QoderWork. Be among the very first to try it out. Can't wait to hear what you build. Stay tuned! 🚀 Token Plan international：qwencloud.com/pricing/token-… China：platform.qianwenai.com/pricing/token-…
English
@beyarkay Seems we got our wish!
x.com/i/status/20783…
Claude@claudeai
Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits. Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit. Demand for Fable has been challenging to predict, which is why we rolled it out to subscription plans in stages, extending access several times as we secured additional capacity.
English
@ZikuD_s Hmm now that's giving the game away, but I mostly just want more accuracy
English
This post showed me the ugly side of X! People were so hateful and I’m attacked about all aspects of my life and career, that my citations and research are fake! There are literal swears thrown at me. 🤦♀️
It stressed me out, I almost wanted to take it down to end the nightmare comments, but I did not want haters to win.
But I’m also thankful for folks who are genuinely kind. Just for my sanity I’ll stick to research posts from now on I guess. 😅
Gowthami@gowthami_s
3000+ citations and level 4 wages in SF, yet H1B rejected. 🤦♀️
English
@ZikuD_s all code is here: github.com/EternalRecursi… but the repo is atrociously structured and I expect you'd be much better off getting fable to one-shot an implementation
English
I developed a method to install and uninstall arbitrary entanglements in models, e.g., making the model love owls only when it becomes emergently misaligned.🧵
English
@steipete Possibly something like this? x.com/ZikuD_s/status…
Hisku@ZikuD_s
Fun little project inspired by autoresearch (by @karpathy ). It spawns standalone agents using Claude/OpenAI/OpenRouter (alternatively cc + codex) in order to decompose, optimize and solve problems against user-defined metrics autonomously.
English
Are we still talking loops or did we shift to graphs yet?
English
Not a huge fan of AI slop but this is brilliant 😂 made my day! Not sure what @elonmusk thinks about it.
Watch Tsar@watch_tsar
One of the best AI videos on IG I've ever seen and had to share with everyone on CT Bunker Boyz by Private Island
Hatfield, East 🇬🇧 English
@vipul_1011 That's a neat trick! Wonder how that pans out for safety, though, given that we monitor CoT to check for model alignment.
English
This is a neat RL trick from Thinking Machines Inkling.
Instead of optimizing only for task success, they optimize:
Reward = Task Reward − λ × (# reasoning tokens)
Then they vary λ across rollouts and pair it with different effort instructions.
The result: model learns that reasoning is a resource to spend, not something to maximize.
English