Galen Hendrix

[Galen Hendrix @greedy_ascent]

@beffjezos @balajis Network States would solve and accelerate so many things. The question is, would they just be snuffed out in the crib at some point via power grabbing from the world powers at be?

[Galen Hendrix @greedy_ascent]

@iamgingertrash @teortaxesTex If one were to relocate to China (e.g. in Shenzhen) as their main AI entrepreneurial hub, how bullish or bearish are you on that?

[Galen Hendrix @greedy_ascent]

@yacineMTB Oooh big claim!
That said, perhaps you’re just restating the narrower claim of pre-LLM ML is actually still useful and not superseded by it?

[Galen Hendrix @greedy_ascent]

@mattpocockuk Lossless verifiable reward

[Galen Hendrix @greedy_ascent]

@mattpocockuk @PhiloGroves Lossless verifiable reward

[Galen Hendrix @greedy_ascent]

@quxiaoyin Being on ground for these, no not true.

Keep in mind that the MLE role already demands general SWE competence.
MLEs are now expected to have even more Senior SWE competence than before, in addition to their MLE competence.

Which actually makes them harder to replace.

[Galen Hendrix @greedy_ascent]

@yacineMTB @jpschroeder Yes, however data can be recovered from weights themselves.
Maybe in the future weights —> ~ total lossy dataset reconstruction can be obtained
(REA for LLMs one can hope 😄)

[Galen Hendrix @greedy_ascent]

@beffjezos Network state

[Galen Hendrix @greedy_ascent]

@GergelyOrosz After not writing any manual lines of code for years now, I am actually moving back to pseudocode, at least for things that actually benefit from good implementation design. Removed all the bloat and dumb small decisions by LLMs

[Galen Hendrix @greedy_ascent]

@FishAudio DRAMA

[Galen Hendrix @greedy_ascent]

@theo On the missing intent, I literally just repeat words / phrases several times if it doesn’t get it! (You see the streaming text happen in realtime so it’s no big deal) The long model pass afterwards cleans it up but usually isn’t even necessary

[Galen Hendrix @greedy_ascent]

@theo Do streaming text in with grok STT API so you can add words here and there as you see them appear + grok does light clean up (all realtime). Then, I have Opus clean up the text as a second pass based on actual context (takes ~10s)

Not good enough for a long precise prompt though

[Galen Hendrix @greedy_ascent]

@__tinygrad__ We need beauty and conciseness rewards.

After vibecoding for years, I’m moving towards pseudocode so I can at least control bloat and intention alignment somewhat, since models these days have become RL fried

[Galen Hendrix @greedy_ascent]

@VictorTaelin Do you hand write bend or vibe code bend? (Is that what Bend in Bend would be for?)

[Galen Hendrix @greedy_ascent]

@Teknium I’m an ex-FAANG MLE, done RL and LLM finetuning + evals at scale, published at ICML+NeurIPS. BUT I’m not in strong interviews.
Do you offer work trials? E.g. I attempt to deliver value in 1 week of free work with minimal to no guidance.
+ Can show hermes apps I built for myself.

[Galen Hendrix @greedy_ascent]

@jparkjmc Lol yeah Amazon will never hold top talent with this BS

[Galen Hendrix @greedy_ascent]

@ThePrimeagen Bro it is just fireworks or tinker APIs and your own labeled data (also easier to get at scale than you think)

[Galen Hendrix @greedy_ascent]

@beffjezos Well, it’s likely just strategic. But yeah we’ll take any bone we can get at this point

[Galen Hendrix @greedy_ascent]

@levelsio Ah no there are ways to make it 100% reliable. Don’t want to leak my alpha there though and then GFW catches on…

[Galen Hendrix @greedy_ascent]

@athcanft Do VR social. Unironically great if you do it right