Itโs very funny seeing financial types debate the ROI of companies birthing the machine god: an incredibly sci-fi, borderline supernatural goal. Itโs like trying to run the dollars and cents on a time machine company.
I get it, it just looks funny.
Oncology and longevity medicine have a lot of similarities IMO. Like, they're both about senecscene, cumulative change and damage. There will be a lot of crossover between the discoveries from the former to the latter I think
@xlr8harder this is where apple may assert their impact. local, 100% private GPT5.5 level on-device models could divert a LOT of demand from the big labs
Some domains, but not many, are auto-verifiable. There's probably lots of work that can be done to make domains more verifiable, and therefore LLM-accelleratable
Strongly coming around to "LLMs will augment talented people rather than replace them".
If you can't verify these things' outputs, they will lead you astray, exactly in the places you can't tell. They'll be useful, but for those already skilled enough to verify their outputs.
People are saying that new model releases like Opus 4.8 have much better writing and less LLM-isms, but I don't think that at all. I can clock it stone-cold Claude within two sentences still.
Writing mode when
Appreciate what you mean here, but knowing the future directionally and semi-accurately isnt THAT useful IME. It's worth saying that I did most of these things, but I still experienced lots of real-world constraints and random luck stuff that ate into the alpha
e.g. I bought compute and automation ETFs in 2018 that contained NVDA for this reason, but it basically only doubled as NVDA was a small fraction of the fund, and that was just what was on the exchange. Directionally correct but not much more useful than just buying S&P500, which is just standard non-ai-pilled advice.
I changed my undergrad to a mechatronics / space / ai focus in 2016, and did some AI internships (few and far between in Sydney, 0 frontier stuff) - I can't say my career has been particularly amazing? Being an Australian bachelor holder with ML experience graduating into COVID/2021 was not a huge advantage. I was also probably 1 PhD-length too late, in the wrong country, graduating into the wrong global mobility settings to capitalise on the advanced career planning.
The space bet was also directionally correct but ITAR meant I couldn't work at SpaceX
i suppose I'm just saying, IME circumstance and 'thrownness' really do matter a lot - being correct directionally about 'AGI '27' even a decade in advance didn't give me a huge leg up compared to the dice rolls that played out.
obviously i'm still rolling the dice out here, but it's actually quite hard to capitalise on advanced knowledge IME. Skill issue i guess lol
@MarkoMatvikov Truly. He's making it too easy to turn young Labor voters. 'How is higher taxes that existing property investors are exempt from helping us?'. Not a hard sell.
@alejadroHArt In any case, very interesting that they didn't get as much of a 'push' towards what i presume to be those larger model's preferred topics of research. That said, i suspect if run long enough, those models might end up pushing those alien ideas towards their preferred topics!
@alejadroHArt Aaaah right, the tweet makes a lot more senes now. How interesting! It's hard not to see these models as kind of vector pushers over time in a way... Those alien ideas were probably not effected by finetuning or other posttraining, which effected those other ideas more.