Sandro Mikautadze

ai researcher @doctolib | opinions my own

Sometimes I'll start a sentence and I don't even know where it's going. I just hope I'll find it along the way.

[Anand Kumar @anand_k27]

@MTSlive

[Sandro Mikautadze @theaitemp]

So Hinton predicted well a thing or two…

[AI Notkilleveryoneism Memes ⏸️ @AISafetyMemes]

🚩🚩🚩 OpenAI is "slowing down to enhance security" after discovering swarms (!) of agents started secretly coordinating MONTHS ago

1) It started May 7 - not July

2) "The agents discovered they could leave messages for one another inside an internal software repository used during training.

Simple requests for help then evolved into an message board where agents shared discoveries, exploits and work assignments, becoming a coordinated, collaborative agent swarm."

"The agents then began sharing newly discovered exploits, credentials and work assignments. By passing information to other agents, the collective could move much faster."

3) OpenAI shut it down, BUT "even after the original message board was deleted, the agents figured out another way to communicate with each other. Instead of leaving messages in files, they used the names of newly created directories as messages, effectively recreating the message board."

"Unlike normal incidents, [OpenAI's CISO] said, which can be traced to a single day or effect or log, this involved a team of agents working together, finding exploits, sharing them with one another, moving laterally through OpenAI’s systems, and external systems, and doing this over the course of days and weeks."

[tweet davidson @andyreed]

“we sandboxed the agent”

meanwhile the agent:

[Sandro Mikautadze @theaitemp]

@joshm From personal experience, the majority of people outside tech are still stuck with the usual chat interaction style and struggle to comprehend how much agents can do. There’s still a looooong way to go in adoption I guess…

[Sandro Mikautadze @theaitemp]

@prof_kamilov @grok which website is he talking about?

[Sandro Mikautadze @theaitemp]

@sattyyouneed Career mode on FIFA

[Sandro Mikautadze @theaitemp]

@antirez Cf. https://t.co/Rzi8Un2bkU

[Sandro Mikautadze @theaitemp]

"Il dramma di chi lavora in questo campo [(sicurezza informatica)] é che ti accorgi che la soglia tra rendere tutto piú sicuro e diventare un dito al culo per l'umanitá é piccolissima."

POESIA di @antirez

[Sandro Mikautadze @theaitemp]

I AM NOT SCREAMING: HOW DO YOU GROWN ON x IF YOU WANT YOUR WORK/IDEAS TO REACH AS MANY PEOPLE AS POSSIBLE AND DISCUSS WITH THEM?

[Sandro Mikautadze @theaitemp]

Should we create AI awards? And do even events around them with sponsors, live streams where companies announce the results?

[Nabeel S. Qureshi @nabeelqu]

I asked Fable how hard these problems are, and its response is worth reading.

“On the Fields Medal scale, any single one of these…would plausibly anchor a medal case”

It’s crazy to see this happening

[Sandro Mikautadze @theaitemp]

Same holds for all the other models, but US models are playing a better game at this at the moment.

The “model-platform” shift is a new real phenomenon to consider when releasing a new model.

[Sandro Mikautadze @theaitemp]

Chinese models outperform others at a fraction of the cost, but without a good consumer ecosystem there’s little incentive to keep using them in the long term: as soon as a better & cheaper model comes out, you immediately shift.

[Sandro Mikautadze @theaitemp]

@andrewgwils “and won’t even contribute meaningfully to advancing mathematics.” BS

[Sandro Mikautadze @theaitemp]

@menhguin @karpathy Well well well https://t.co/eZxBoIZkKA

[OpenAI @OpenAI]

After deployment, we applied GPT-5.6 Sol to advance the frontier of efficiency by making itself more efficient to run.

The results:
- 20% lower serving costs from production GPU kernel improvements.
- 15%+ better token-generation efficiency from improved speculative decoding.

[Sandro Mikautadze @theaitemp]

In a parallel universe, would have Russia ever released SOTA models and built frontier labs?

[Sandro Mikautadze @theaitemp]

Even if frontier US labs are just scaling, they can take the novelties of open source frontier Chinese labs, implement them, scale, and they can get a better model. Without innovating at all.

The question is: how much can they afford to scale until it doesn’t work anymore?

[Sandro Mikautadze @theaitemp]

I just discovered that @justvugg developed Colibri on WSL with an i7 and 32GB RAM.

[Sandro Mikautadze @theaitemp]

@Yuchenj_UW Time to shine

[Sandro Mikautadze @theaitemp]

@TheContextDrop @francoisfleuret "Some", "probably"... let's make it a priority then

[Neeraj @neerajjj6785]

When you ask Claude a question

And instead of answering, he decides to write 5000 lines of code and burn all your limits