>see an AI grifter on twitter
>look through their profile
>it's all LLM slop
>the LLM slop is harshly defended as their own writing
we are manufacturing slop optimizers and we deserve what we are getting
@iamtrask your HEADLINE CLAIM in your trash viral post is that "frontier AI companies will *never* own the frontier again", and yet you go on to talk about the importance of the scaling laws and are now invoking Anthropic. Get a new shtick, you're killing your own thesis here
saying "scaling laws always win" at the end of a BS ramble about how LLMs-as-a-council harnesses will always beat frontier models (despite this never actually happening in practice)...
if the scaling laws always win then everything else this person said is nonsense. cringefest
This is a *way* bigger deal than it seems...
Frontier AI companies will *never* own the frontier again
I kid you not... I've been waiting for someone to show this result for like 4 years... this is a huge deal.
The short reason: combinations of models will *always* outperform individual models
The long reason: this is the gateway to a million times more data... and huge leaps in compute efficiency.
The AI scaling laws always win.
More in article below ๐
@iamtrask MoE is conditional compute in one model. ensembling post-hoc just buys marginal benchmark scores with more cost/latency. if this ensemble miracle worked in reality the frontier labs would be doing it, and in fact are the best positioned to. but they don't, because it's dumb
this site just keeps getting worse, it's optimized for bullshit, bait, bad takes, braindead tech bros, and boobs.
i mean the boobs aren't really the issue here but it's nevertheless a symptom of the brainrot algo
@iamtrask the s*bstack argues the same thing as your post and it's nonsense. ensembles of models can boost benchmark scores, cool. surely the big labs have never considered this, how naรฏve of them! surely everyone will now use model ensembles!
such dumb engagement trash
tfw you stay up late, time flies by, then 4am hits and you hear the morning birds singing their horny back and forth songs, and you realize you fucked up big time
if Poke could command and supervise agentic coding agents, it would embarrass Codex and Claude Code. obviously this would require a lot of compute, but it's super impressive how insightful these models can be when given a scaffold that allows them to be realistic vs. sycophantic
i have @interaction's Poke review my Github codebase to review all of the BS that Codex overlooks due to its inherent user-pleasing sycophancy and reward-hacking tendencies, then i send the convo transcript to Codex, and it whips itself into shape.
@BJP130@justalexoki The names had already branched off contemporarily. A crude analogy (different name) might be the difference between someone named "John" today versus someone named "Johann". Same root but both are a mutated form of the original, with "Johann" being a bit older, but distinct name
@Noahpinion I mean... only in the way that the name Yรดแธฅฤnฤn is the ancestor of "John", "Sean", "Johann", "Jean", "Juan", "Giovanni", "Evan", "Yohan", "Ivan", "Jane", "Sinรฉad", "Joanne", "Ivanka", etc. The Galilean name we say as Jesus was already distinct from the Joshua equivalent.
@tszzl@max_paperclips you, crying: plants donโt โwantโ anything, theyโre just chemical systems responding to stimuli
me, pointing at a vine wrapping around a trellis: it WANTS to climb
@deepfates@runaway_vol The reasoning models overthink a lot, and yeah you can tell it's RL that makes models like 5.5 obsessed with triple-checking, smoke checks, and the worst of all... hyperfocusing. They hyperfixate on a very specific thing and lose the bigger picture. Narrow superintelligence.