@GaryMarcus@CristianFl46703 Idk Gary I could have sworn you said LLM reasoning was "cooked". Is the "knockout blow" to LLMs here yet ? You're bringing up the same complaints about the Unit Distance Conjecture result only for Noam Brown to confirm it was 1. a general purpose LLM and 2. didn't use lean.
@Mefa_MC@ilyasut@OpenAI Bro's not aware of the coup attempt lore against Sam and Ilya having an existential crisis deciding OAI was a dangerous place to continue working on this tech.
@GaryMarcus Gary the models are already vastly better than the majority of humans with zero need for tool calling. No neurosymbolic augmentation needed *unlike what you've been saying for years
@DKokotajlo@AIExplainedYT@thlarsen@romeovdean@eli_lifland I know you had said previously you thought the AI 2027 timelines were slightly too aggressive ie. (a few months ) is that still the case or have you defaulted back to the timeline laid out in AI2027?
@yacineMTB I see the OAI propaganda machine is in full swing. I'm not all in on Anthropic and many of the ways they have approached particular issues is troubling but at least it's not "start a bidding war between Russia , China and the US for control of humanities future".
@ZixuanLi_ CAD design would be huge , looks like some of the other frontier labs are beginning on it . But being able to bring the AI gains in software to hardware would be awesome!
@goodworse@xatacrypt It by no means destroyed everyone it didn't top the benchmarks in really any metric and Meta continuing to use Claude even after it dropped is telling in terms of it's actual utility. It's not a bad model but it wasn't SoTA , made even less impressive by how GPU rich meta is.
@TheZvi I'm not sure adding KYC to customers is enough to actually satisfy the government if we assume access to US citizens only is genuinely a deal or no deal thing. If a non-US person uses a friends account is Anthropic on the hook ? How would Anthropic even police that? Ect...
@5_utr@joegibbs98@InverseMarcus "below human level" , as long as you ignore chess, go , StarCraft , Gran Turismo, competition programming , math competitions , protein structure prediction , imagenet classification, facial recognition , skin melano...........
@EMostaque downweighting --> amount of raw compute needed for the frontier. | Nvidias chip dominance. Western lab talent gap.
Upweighting ----> Chinese research talent | how cracked Huawei chips are | CCPs amount of autonomy they give frontier labs
@GaryMarcus@Sam_AGI_Vietnam Fwiw Hassabis was continually late to the party on the LLM train because of his priors. And seeing that GDM is trying to pivot to world models it shows he still doesn't have as much faith as the other labs do.
@thealpharaccoon@soundboy@ESYudkowsky Just to pull two counterexamples to this narrative, Stuart Russell and Max Tegmark are both respected researchers who have been voicing concerns about this technology for years (decades) , I don't think they're in it for the money or coping that they can't compete with OpenAI.
@momentofdeep @thiccythot_ Y axis should be portfolio growth rate , X axis should be % of portfolio allocated , the graph should be linear if you're actually plotting EV vs size (ignoring slippage/market impact). Both of these graphs are kinda jank imo.
I want to give away a free 1 year subscription to my paid newsletter on exchanges & market structure (a $249 value). To enter:
- Follow @HideNotSlide
- Like & RT this tweet
- Be thankful for today!
Will choose a winner in 48 hours.
https://t.co/rsfAT80cW9