lilpuke

@_hughhan @gdb so.. token futures

@gdb oai is incentivized to do this because token costs go down as open source models and harnesses catch up. the bottleneck is compute, not the model. with this thesis, why would one commit 1-3 years to oai as opposed to a compute provider?

we are offering discounted tokens and certainty on capacity availability in exchange for 1-3 year commits. we expect that the world will feel increasingly capacity constrained for the next while, as models continue to get much more useful.

@blc_16 Noob

MIT just released a new RL method called Pedagogical RL. The main lesson -> correct reasoning traces can still be bad training data. It is a similar concept to teaching someone backprop. Say you have a tiny computation graph: z = wx + b a = ReLU(z) L = (a - y)² If you already understand backprop, you can jump straight to the gradient: dL/dw = 2(a - y) · 1[z > 0] · x The answer is correct but it skips the reasoning process. To get there, you need to break the computation into local pieces: dL/da = 2(a - y) da/dz = 1[z > 0] dz/dw = x Then backprop is just composing those local derivatives backward through the graph: dL/dw = dL/da · da/dz · dz/dw = 2(a - y) · 1[z > 0] · x Showing a student the final gradient does not teach them how to find gradients on new graphs. Even telling them “just use the chain rule” may be too large of a jump if they do not understand how to decompose the computation into intermediate nodes and local derivatives. Reasoning RL has the same failure mode. A rollout can pass the verifier while containing one step the student model basically never would have taken. The trajectory gets the answer right, but the learning signal is brittle because the path is too far from the student’s current policy. Pedagogical RL trains a privileged teacher that knows the answer, then rewards it for producing trajectories that stay learnable for the student. The trick is to use a spike-aware reward. It penalizes single huge surprise gaps in the trajectory, even when the average likelihood of the trajectory looks fine. Then the student learns with surprisal-gated imitation, where teacher tokens that are still too surprising get downweighted. The teacher is learning how to teach at the student’s current level. Pedagogical RL makes RL more efficient by efficiently selecting trajectories the student is most ready to learn from. Less waiting for the model to get lucky rollouts. More training signal from examples that meet the student where it is. Full blog in comments

@_hughhan LGTM

hey dude can you review this pr real quick

@_hughhan @RhysSullivan Good choice for a good piece of technology

@RhysSullivan for gestaltd.ai, we chose the indexeddb protocol. it's been working quite well for us

what are the best pluggable storage implementations you've seen that let people bring their own storage to your code? figuring out the public sdk shape for Executor, the simple path of static sources loading is straight forward, but dynamic tool loading requires state

@ericwallerstein @zkriv96 Same

@zkriv96 awaiting the inaugural x article or substack post from mr cfa. bless us with your knowledge

There’s a narrative getting louder that China is the winner from this conflict as nations are now more incentivized to pursue renewables + China sells the cheapest clean tech. Can at least see that being the case for nations that can’t afford SPRs

@zkriv96 Agreed

@_hughhan Caffeine keeps Hugh going too

macos caffeine + hotspot on the commutes to work to keep the local agents running

@0xnirmal Woah

who we picking here

@pavitarsaini I want this

I made a Chrome extension that turns localhost into a visual editor for Cursor. Click any element on your dev site → describe what you want changed → it automatically sends the edit request to Cursor in the background with the element context. Here's how it works...

@Askronniegg Needed this

I MADE AN ENHANCE SHAMAN PVP GUIDE. #1 SHUFF World & 3v3 Rank 1 MAX SIMPLE. GO ONE SHOT NERDS: youtu.be/sX-Xt9fIZ9Y

@iScienceLuvr I can’t tell if this is satire

This paper claims LLMs are better at selecting successful founders than VCs "We introduce VCBench, the first benchmark for predicting founder success in venture capital (VC)" "most models surpass human benchmarks"

Typescript is my love language

Claude is such a softie.. it told me to pack tissues for the wedding I’m going to in case I cry

@omkvr Literally I don’t get it

who the fuck is the target audience for the iphone air lmao like what

Switching out my usual tech-doomerism with tech-optimism from now on. A premonition of the future: The next CEO of Tesla will be a dolphin who will communicate through a human-to-dolphin fine-tuned LLM translator and build underwater autonomous Uber submarines to corner the oceanic travel market.

@0xnirmal 👀

I feel so seen

AI can solve everything except for task management software. The heat death of the universe will come before the final evolution of linear.

More thoughts on why big tech has run out of work: - all new work is iterative and small - all frameworks are mature i.e. react isn’t being built again today. Databases are solved. - everything that needed to be built over the past decade was built. I.e. hiring website.. done and they aren’t rebuilding anymore like they did back in 2010s for on prem -> cloud - ai code is small. Not a lot of surface area - base models > fine tuned models - .. every ml team that had build custom “old” ml pipelines is gone and using ai in place

There is not enough work to go around at big tech in 2025. This is not sustainable.