Ilya Venger

Opinions are my own. AI, Cyber and Us.

[Ilya Venger @ivenger]

@kurtsaltrichter 7 times what Americans spend on phone, streaming, and internet combined --> sounds not that huge.

[Ilya Venger @ivenger]

I managed to crash my Mac. Achievement unlocked

[Ilya Venger @ivenger]

@emollick It stopped being a model of human language after that long a time in various RL environments.

[Ilya Venger @ivenger]

@AndrewPrifer @OpenAIDevs Yes...

[Ilya Venger @ivenger]

@dpetrou @ylecun @Michael_J_Black @SimonGoodman_ Not really. GPT-3.5 didn't hallucinate to the same extent as Galactica (Galactica wasn't RLHF'ed) and was not positioned as a scientific tool which raises the bar on consistency and truthfulness. And ChatGPT also had a decent accessible interface.

[Ilya Venger @ivenger]

@ylecun @Michael_J_Black @SimonGoodman_ If only Galactica was RLHFed

[Ilya Venger @ivenger]

This shows why it's important to close the experimental loop in the physical world.

[Anthropic @AnthropicAI]

Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages. Beside the enzyme’s gene sits a long array of repeating DNA—a structure that looks somewhat similar to CRISPR.

We don’t yet understand what this system does, but only a handful of known systems share its features, and all of them are able to cut, copy, and paste DNA. Historically, the discovery of such programmable systems has helped revolutionize medicine. CRISPR, for instance, is now the foundation of genetic medicines. But it will take much more work to learn what this system does, and whether it can be put to similar use.

Read more: https://t.co/RuEosScSMb

[Ilya Venger @ivenger]

@eyesofastorm @ezraklein @JensenHuang He had the same attitude with Dwarkesh. Basically, his position is that Nvidia should win never mind everything else. He's not a loser!

[Ilya Venger @ivenger]

"These ideas have been around for a long time. It's software..." - yep - but These ideas were science fiction and futuristic until very recently.

[Ezra Klein @ezraklein]

Tomorrow on the show: @JensenHuang, the CEO of NVIDIA, who thinks A.I. fear is getting way out of hand.

[Ilya Venger @ivenger]

@yoavgo That is to say the finding might be interesting for future work, but it's not proven.
A little but similar to Mythos (not quite there yet). In cyber the models are able to find an exploit. But it's not yet much better than a skilled human hacker.

[Ilya Venger @ivenger]

@yoavgo As a PhD in systems biology and bioinformatics (albeit a long time ago) - an interesting finding. Something a PhD student or a bright masters might achieve in 2 years.
It's a start, but there's no loop yet - the model hasn't learned or adjusted the harness from an experiment.

[Ilya Venger @ivenger]

@kimmonismus 4/n
"Patch's propositions were therefore poohpoohed in official USIB pronouncements and by the White House spokesman, and there was also a campaign to discredit Patch. His claims were equated with the many irrational fears and prejudices which had arisen in society at the time"

[Ilya Venger @ivenger]

@kimmonismus 3/n
"... Professor Eldon Patch presented a paper in which he maintained that, even when impregnated in the manner described above, a computer may cross the so-called "axiological threshold" and question every principle instilled in it."

[Ilya Venger @ivenger]

@bratrat @8teAPi @Indian_Bronson I recently reread the "Foreword" to Golem XIV. It is uncannily prophetic. RSI, issues of alignment and people dunking on "doomers" all got mentioned in the first three pages. 1973 mind you...

[Ilya Venger @ivenger]

@8teAPi Mathematicians are the first to find themselves wandering the halls of The Library of Babel.

[Ilya Venger @ivenger]

@kimmonismus So now we're just arguing on semantics rather than the path.
"We must pair the frontier" means we must advance in lockstep with our ability to do so safely.

[Ilya Venger @ivenger]

I've been playing with Jev, but then ended up calibrating the benchmark instead of the model.
I took the AG News dataset - oldie but goodie. Jev showed not great results - ECE of 0.08; but then I looked at the errors, and some data labels were off... benchmarks are difficult