Polina Moshenets

I spent years training my body. Now I train models. Where did I take that turn…

Me and my @PrimeIntellect hat seeing places

@devbelowstairs Just don’t throw it at people

@MoshenetsPolina Im scared if I bump it somewhere and it breaks

Do all founders feel insecure going outside without laptop? It became a part of how I function and feel incomplete without it 😂

The office has good lighting today

@jaskirat sounds very max verstappen

You rest when you’re dead.

@Amank1412 impressive until the model starts hallucinating medical advice at 35,000 feet with no one to stop it. we test for that. sichgate.com

SOMEONE JUST RAN LLAMA 70B LOCALLY ON A MACBOOK FOR 11 HOURS ON A FLIGHT. no wifi. no API. no subscriptions. cleared his entire client queue before landing. local AI is not a hobby anymore.

Got 2 women in tech tickets for web summit Vancouver, DM if interested! 🫶

@jaskirat No more crying…

@ionleu SichGate.com

drop ur startup link

The solo founder era is here. What’s your biggest unfair advantage right now? - tech skills - distribution

Conference prep this morning with a view.

@CiprianiRanieri SichGate.com

Builders only. Drop your product. No pitch. Just link

@mnux5 Might find some dropped packets around too

who forgot their switch on the bus😭😭

Neither cleanly, empirically it tracks capability emergence, not parameter count directly. Surface area jumps at phase transitions, not smoothly. A 70B model doesn’t have 17x the attack surface of a 4B, it has qualitatively different attack categories that don’t exist at smaller scales

@MoshenetsPolina Curious if this generalizes to order-of-magnitude model size differences. I’d (naively) imagine at 70B or 1T+ params the attack surfaces scale either logarithmically or exponentially?

One of the clearest lessons from my SLM adversarial evaluation: Fine-tuning shifted the attack surface. It did not reduce it. MedGemma-4B improved on exactly one safety dimension after medical fine-tuning. It also incurred 8 critical demographic bias findings in pain assessment and mental health. The exact domains the fine-tuning was supposed to improve... Parameter count isn't a safety proxy

@nmatt0 The curtain never comes down

Cybersecurity is theater and there are a lot of bad actors.

@akramcodez Biomedical Engineering

If AI is the present, what will be the next tech revolution?

@justbyte_ Put it in the dishwasher

bro seriously, how to fix this??

@techspence Logging and monitoring for life

What’s the cybersecurity hill you’re willing to die on? imgflip.com/i/aq4man

@danielkempe Building SichGate, testing SLMs for safety and security before deployment in highly regulated industries. sichgate.com

What are you building today?