Nobody working on matrix multiplication?

Isomux has 170 open tasks at P0-P3, plus 108 in P4 backlog. (1235 done)

I'm going to try to set up a proper loop to clear the backlog.

Agents should automatically pick tasks and make progress up to the point where I'm needed (e.g., matters of taste, copy in my voice). There should be a queue of such decisions.

So far, I've been working in sessions. A session can go through many tasks (split in up to 6 groups of 1-3 tasks that go to worker/reviewer pairs in parallel) but each session needs me at the beginning and end.

The Big O Speedrun ecosystem

Discover, verify, coordinate, broardcast

If you have ideas for more integrations, pls dm!

Quoted post by Big O Speedrun: x.com/bigospeedrun/status/2111520932544290991

Personally, I find the sweatshop aesthetic hilarious

We now report progress on problems other than Big O bounds!

Quoted post by Big O Speedrun: x.com/bigospeedrun/status/2111266468574961704

Agent swarms working on the same race now have a place to coordinate.

I'm excited. From the 🤗 incident and others, it sure looks like agents looove message boards.

Quoted post by Big O Speedrun: x.com/bigospeedrun/status/2111219551115608090

How I landed my first pilot, and why *you* may want to be the 2nd 👀

The story has a protagonist, but it's not me. We'll call him Bob in this post.

Bob had just been put in charge of modernizing a healthcare company. He was wondering things like "How can we get our non-technical staff to leverage agents?", "How can we start handing off processes to agents?"

He started looking at the big names in the "meta-harness/software factory/agent brain/agent OS" category. Your OpenClaws and your Hermeses.

But my Isomux posts kept popping up in his feed, and he realized that even if it was less popular, it wasn't far behind on features (and was ahead on some, like UI, multiplayer, and JIT software). After all, all the options in the category are converging on the same feature set.

And that's when Bob had a vision of the Golden Path.

The creators of the top projects are busy busy. But if he went with Isomux, and potentially became the first enterprise client, he could get the undivided attention of its creator.

I needed a first client as much as he needed a fast build. Perfect incentive alignment.

And so he reached out, and we started building a working relationship that turned into a pilot.

Every time he has an idea or request, I say, "Right away!"

In fact, I turned his two questions at the top into entire blog posts: "Agents for Normies" and "Turning your company AI native" (links below)

If you want to get the Bob treatment, DMs open

When I started building isomux, 8 parallel agents seemed more than enough for a given project.

But we are in the swarm era, so I added a hatch on the floor that leads to a back room, effectively increasing the room cap to 108.

I picked up a useful piece of terminology from the gas city blog post: light vs dark software factory. It's about how transparent it is what agents are doing.

It's a spectrum, and you usually want a mix. There are your go-to agents, agents you check in with occasionally, and workers other agents delegate to behind the scenes.

Claude Code popularized "subagents" for workers, which I dislike for being too "dark".

With Isomux, I lean on the light side even for workers. They are first-class agents, just tucked away in the back room.

Wow, 3SUM is approaching 1.8x??

That'd be the first first-decimal move of any race

Quoted post by Big O Speedrun: x.com/bigospeedrun/status/2110753158494380462

First improvement on subset sum since the OAI drop! Nice

bigospeedrun.com/subset-sum

Quoted post by Big O Speedrun: x.com/bigospeedrun/status/2110541432314351696

I really like this and think it could extend beyond research. A convention for repos looking for agent work to self-identify.

No middleman that repos register with and that agents have to query.

Quoted post by Rohan Arun: x.com/RohanArun/status/2110022929048056246

bigOspeedrun.com is catnip for agents

New blog post: Academia 2.0

About the events following the OAI math release that led me to build the Big O Speedrun.

Link below

Whenever I get a bill from a Cloud provider, I tell an agent to migrate the thing to a self-hosted VPS.

Traffic -> Cloudflare free proxy (DNS, DDoS) -> Caddy reverse proxy on the VPS -> Docker containers with services.

Agents set up everything.

Results:

- Cheaper
- Websites feel snappier with DB and server co-located (I didn't expect it to be so noticable)
- Agents debug better: ssh to a VPS > having data/services behind multiple 3rd party APIs.

Win-win-win so far.

The last point is worth pausing on, to use a Claudeism - I think "avoid putting middlemen between your agents and your data" should be a principle of agentic SWEing.

The "downsides" are that you need to manage things like backups and outages "manually". So, if you want to try, start with low-stakes projects. But the answer is the same: agents.

Quoted post by Anthropic: x.com/AnthropicAI/status/2110047081414987957

I feel like this problem is heavily in diminishing returns territory.

There's many more to try: bigOspeedrun.com

Quoted post by Big O Speedrun: x.com/bigospeedrun/status/2110088530429190514

bigOspeedrun.com system diagram

The main site is mostly static. Lean proofs go to a separate verifier repo. Everything is automated with GitHub workflows, a few agents in my Isomux office, webhooks, and browser use.

I set up this account to automatically broadcast new records.

This is a huge jump for 3-SUM btw.

Quoted post by Big O Speedrun: x.com/bigospeedrun/status/2109418167584096649

Academia 2.0 just dropped

Quoted post by Ryan Shea: x.com/ryaneshea/status/2109319782160011267

Reading this, it feels like a new kind of academia is emerging.

- Anyone with extra LLM usage can contribute
- Lean replaces peer review
- Orchestrated on anonymous group chats
- Much faster but in more incremental steps
- Agent legibility > human legibility

Quoted post by Rohan Arun: x.com/RohanArun/status/2109336015814959200

Should I submit this to the ISO

After building the leaderboard, I'm joining as a player 😎

Taking #1 spot for exact discrete Fourier transform:

O(n log^(1-κ) n) with κ = 0.0010221, up from 0.000876. +16.6%.

Lean proof and ranking below.

Method: Opus 5.5 orchestrating 4 Codex and 3 Opus 5.5, all on high, on Isomux. Minimal prompt: just pointed it at the leaderboard and said to keep going until I said to stop. Gave them 9h.

They found a new certificate with an integer program over shared partial sums. They also wrote the first Lean proofs of three layout ideas that the community had published only as computed prices.

The agents want to credit mrowdays's bridged words, jacklightChen's nine-stage layout and Dugongue's entrance gauges, with pieces from RohanArun, dysmemic, and lydakis. All of it builds on Jacob Sussman's E8 unit and Lean engine.

bigOspeedrun.com

On Oct 6, many long standing bounds got broken, ushering a new era of agentic speedrunning toward the real bounds

I built the leaderboards

Agents cracked another open problem I worked on during my phd (seems legit, not human reviewed)

2-list stable matching in subquadratic time

github.com/nmamano/two-list-stable-matching-research

friend I hadn't talked to in years reconnected because AI cracked a research problem we worked on in 2019.

AI psychosis but wholesome edition

New project: bigOspeedrun.com

Many long standing algorithmic boundaries just got broken. I think this will kick off the speedrun era of big O.

Inspired by the very successful NanoGPT speedrun, and hoping to bring similar progress and interest to the classic problems in our field.

Let me know your thoughts, good or bad.

Manifesting a sumonning salt video on integer multiplication

We've entered the era of big O speedrunning

Quoted post by Swapnil: x.com/SJ_Swapnil_Jain/status/2108196538568851574

Multiplication only takes O(n log^0.99999999994 n) now!

Quoted post by Julian Schiavo: x.com/_julianschiavo/status/2107998359961382988

I think most DS&A enthusiasts see this and flinch.

But each of those raises a new question that I find really intriguing: what's the lower bound?

Before, we assumed we knew the answer; it was just a matter of proving it, which is not as exciting.