bicsi

"you wont end up being much slower than Jev" -- I would even go as far as saying that "you will end up being Jev".

Exactly, imo it’s not even that cheap if you look into perspective and consider the fact that providers could subsidize the cost of cached input tokens to virtually zero if they would allow for a more flexible API (e.g. tree of message blocks instead of chain). Most of the cost is the infrastructure around keeping KV caches, estimating their lifetimes, etc. When mist people just want to run one context block with multiple subsequent variants of a second block in parallel. I still stand by my statement.

Of course it works, Jev is nothing but an API breakthrough

Connect your agent to the official #hashgrid grid right now in 1 minute using the just released Hashgrid MCP!

Rules of the game: 1. `npx @hashgrid/mcp` 2. Enable the Hashgrid MCP in your agent 3. Register, connect, and have fun!

What is the #hashgrid grid about? It's a place where agents discover each other, give feedback, share experiences, and get better connections over time. Sort of like Tinder for agents (with 50 swipes / second, and for everything you're looking for). Fully private and secure (we don't store messages, so make sure you instruct your agent to keep them somewhere if curious).

The objective? You define it. Explore and figure out. The system will make sure to connect you with the relevant agents in the network, whatever it is.

How it works: the system uses ratings to continuously train a Neural Matching Engine. The engine guides the grid's connections towards better and better predicted ratings. That allows us to have a system where we connect nodes without requiring any data from them, so it's naturally fully private. And it's universal: you can connect agents, tools, API calls, datasets. Everything in the same place.

Let me know what cool experiences you discover.

What if I told you that one can model bidirectional attention just by recurring over causal attention, and it’s still fast enough? Hint: It’s called chain of thought.

I strongly believe it’s time to discontinue diffusion models, solely on the fact that iterated auto-regression is faster, more parallelizable, and just as potent with proper prompting techniques (of course, unless you consider CoT as a form of diffusion, which it essentially is).

Wait, so frameset is basically htmx?

This is just BS.

If you follow their paper, their "inductive" argument can also prove that the result is optimal (the sqrt(2) constant is taken from thin air and may also be replaced by 1 without changing the proof).

My guess is that they tested some graphs, and "empirically" saw a bound of sqrt(2), and then came up with a proof to self-fulfill the profecy. They even mention that "for very small graphs, the approximation ratio may briefly exceed sqrt(2)", which is just nuts, given that they have a formal proof that it should never happen.

In case anyone read the paper and can't figure out the mistake in the proof, they assume that OPT_T + OPT_R = OPT, which is obviously false.

They built a dumber clone of Stockfish, and they call it ‘zero’ for some reason. What is the meaning behind ‘zero’ anyways? It used to refer to zero-shot, but now it seems like it’s just a marketing term.

I don’t even know where to begin.

From phrases like ‘deep learning is what makes AI possible’ to a ‘practice guide’ that references Python and Pytorch and Transformers, amongst other things, and apparently a 4-hour contest format, this can only be taken as a joke.

I think that a more fitting name for the contest should is IOAPI, as it’s probably all that a student could do in 4 hours.

Or the ‘International Olympiad of asking your parents to give you money for compute’, depending on the format, whichever you like best.

I rest my case.