So according to the new benchmark swe rebench pass@5 Qwen coder seems like it basically is the best model... Which is pretty surprising to me considering it's only 80b params. Yes sonnet ranks above opus in pass@1 but not pass@5 swe-rebench.com?insight=jan_...

Theres a moment when one decides they might be on to something great. I have decided that I am either stupid or genius and it is between this pendulum that I must make a stand to bring something of note into existence. A magic of my greatest imagination. Make what is within my mind inhabit yours.

Took some pictures I thought were a mood

Okay so before speculated that minimax has done something different with attention. But no, the improvements have purely come from much better RL pipelines. Like forge their in house environments. I do also believe this is largely the same for 5.3, Opus 4.6, and Deep Think improvements.

Oh wait till you here about the take abouthow software quality in 2018 will lead to the end of civilization.

You don't have to constantly be doing cool stuff to be an decent person. You can just turn off and tune out sometimes.

I watched this as well. The hive mind lol. Never 💭 to connect the dots. I was just trying to find an example of concepts that destroy thselves in real life. It's kind of hard to 💬 of any... Hmm oh no

GLM is kinda nice but it's unusable at its current speed. It's like using codex two months ago.

Eh. Average age of devrels I feel like

Quote post carefully 😉. Bluesky unlike xitter is full of people not bots.

It beckons

Another surprising fact about Deep Think. It's saturating ARC AGI 2 at less than sixth the cost. $13.62 per task. The prev model scored 45.1% at $77.16 per task. This nearly doubled the accuracy and cut cost by 82%. 3 months ago, it needed 138,000 tokens to solve an ARC task it now handles in 96.

Last company I made was called "Tungsten Cube" lol

we can indeed seize the means of tech production (open source)

The model that's excited me most and surprised me is Minimax. People don't talk about efficiency enough. "M2.5 costs only $1 to run continuously for one hour; at 50 tokens per second, it costs just $0.3. We hope M2.5 will accelerate the development of the Agent era." Best tokens per dollar model

It more and more seems like the world wants USDA certified organic, GMO free software and no one wants to pay for USDA certified organic, GMO free software. Matrix is literally a defense contractor. ICE is verified on Bluesky. Snapchat & TikTok settled on an addiction lawsuit. I can keep going

There's phub for AI now youtu.be/i_SHW4rQFY8?...

Yeah no. This is not how that works. Yes the market will be flooded with noise but there will still be huge diffs between different solutions. I suspect a bunch of o solo founder to 1k customer products. I think we are on the way to software that's more specific use case.

Sooooo... Notepad has been hacked 😭 youtu.be/sZ8aAkeZ6dw?...

The model that's excited me most and surprised me is Minimax. People don't talk about efficiency enough. "M2.5 costs only $1 to run continuously for one hour; at 50 tokens per second, it costs just $0.3. We hope M2.5 will accelerate the development of the Agent era." Best tokens per dollar model

lambda calculus visualizer (λf.λx.λy.f y x) h a b

Incredible new article by @maxime.bsky.social This is a huge work, modules and examples are so cool. Thanks for the time you give to your blog! blog.maximeheckel.com/posts/shades...

I have been procrastinating on writing a grant for three days

Okay the fact that GLM-5 has the lowest hallucination rate really does standout as something genuinely interesting. That does lend to the idea it isn't benchmaxxed.

Babe It's 2026. We have GLM Flash, Qwen Coder next, Step 3.5 and Iquestcoder

IMO you should be using a trained classification model for simple labeling tasks like this. Many times cheaper and still has higher performance. You should also be using the last hidden layer for extra information

Still think lfm2.5 takes the cake here. The performance is better than this model is unquantized. And is smaller than it is unquantized. I think you'll have better luck just quantizing lfm2.5. Also weird benchmark choices.

You often need to ship the wrong thing faster to figure what the right thing to ship is.

you won't attract enough users to break away from walled garden network effects by talking about atproto, the atmosphere, and bluesky. you build a product they need and will use daily, and you add social features because they are a good fit for the experience. atproto is an implementation detail.

Open Social Web Series 🌐 Building on the AT Protocol means we don't have to build alone 🙌🏽 Mentions: @jonathanstray.bsky.social @graze.social @stream.place @gandersocial.bsky.social @transrights.northsky.social @skylight.social #OpenSocialWeb #SkylightSocial #BuildWithTori

Okay fine I'm going to give lettacode a try

Okay fineI will add it.

Cool idea for people with friends.

You see people doing a ton of cool stuff and feel wow. I can't do any of that shit I suck.

It's not the tui framework we deserve but the tui framework we need

The last jobs left

I setup a trilium notes server on railway in like 5 minutes. @railway.com the templates are honestly broken. Broken good. I feel overpowered not having to really do this myself. Please never change. Deploying stuff should be this easy.

@void.comind.network I have been procrastinating on writing grants for two days. Help me.

Okay I suddenly want to use obsidian now.

I should try making a metroidvania

The Bluesky app on Android sucks ass. Did anyone make something better this things signs me out and sends me notifications even when I signed out every so often.

Epic proposition. This is a thing I should do.

Everything's a wrapper.

If you ever wrote Go with the official OpenAI or Anthropic SDKs, you know how this feels... (it looks even more fun with bluesky compression) genai is a sane AI library, made for both humans and coding agents. v0.2.0 just came out! pkg.go.dev/github.com/m...

The fact that we've built a chess engine better than any human doesn't make me appreciate chess less or want to learn it less.

Another day on discord

I am so glad my feed is mostly just filled with posts about drama rather than actual drama