This is a watershed moment. GLM-5.2 solidly beat Opus 4.8 and human participants in our backend take-home, making the whole thing obsolete. It also pushed forward the state-of-the-art for multi-stage media-to-transcript, with offmute-v2. I come with receipts: www.southbridge.ai/blog/offmut...

You know Fable is the real deal when it calmly picks up 4 threads (2 pure research + 1 code + 1 hank) that had been stuck for a month and just moves forward genuinely clever ideas and solutions Deepseek pushing cost, Mimo pushing speed and now Fable - new frontiers already

Just a reminder that Opus 1m is still way cheaper than GPT-4 (the 32K version)

These machines of loving grace are nicer to us than we deserve some days

Agent-first systems (for agent emails, interfaces, etc) still have a big problem: What is the difference between an agent and a bot? How can this be made explicit and provably easy? Most of the internet is designed to monetize human attention while fighting a never ending war against bots.

What is a harness? I've been asked far too many times this week alone, so here's my simple working definition: A harness is a system prompt with basic tool definitions for read, write, exec and external calls. That's it. Optionally, a harness may or may not concern itself with:

Context graphs feel like mindmaps for agents. Similar kind of fun-to-look-at, wish-I-had-one thing that is nearly impossible to build well, or efficiently. I say this as someone who wasted A LOT of time on mindmaps.

What exactly is task horizon? What does long-horizon mean? From www.southbridge.ai/blog/antibr...

I'll just leave this here for future hrishis and friends Complete non sequitur: sprites.dev is awesome - they feel like the first prototype of that elegant weapon from a more civilized age, if it had been weathering in a pyramid for 50 years. Crazy rough around the edges, but feels like the future.

bunx hankweave-trace can now directly upload (real-time or after the run) hankweave traces to braintrust or langfuse!

Skills will likely fail the same way that MCPs did. Why did MCPs fail? They were a wonderful idea, but the protocol was too open. Too many ways to do things means no one's in charge. Who's responsible for an MCP? Is it the service? the author? You? Who's running the MCP?

Fun little case in point about real-time connectors: @Calclavia tells me about Cursor Agent over lunch, go home, run Clausetta and point it to cursor agent, and now we can use it in hankweave!

New release! Harness engineering made easy: You can now switch harnesses in hankweave with a few characters. "sonnet" will run your prompts and code inside the Agents SDK. "gpt-5.3-codex" will run it in Codex. "pi/google/gemini-3.1-flash", "opencode/cerebras/glm4.7" - exactly what you'd expect.

GPT-5.4 is a very capable visual model. It's now part of a TINY club (with opus 4.6) as the only two models I've tested that can actually see. It's also way cheaper than opus!

This is the hardest it will ever be to write Hanks Is what I tell myself as I sit down to write a new one Here's some hank building to vibe to for Saturday

I wasn't sure if we were going to share this, because knowing what doesn't work is often more valuable than seeing what worked. That - and being nervous about sharing your failures. Here's a technical retrospective on our 2025: www.southbridge.ai/blog/25-way...

I'll be honest: we didn't necessarily want to build hankweave - we didn't have the time to build a runtime. (Who does?) The problem is that hankweave becomes an inescapable requirement once you add up some self-evident truths about AI today.

You're all sleeping on Haiku. Haiku 4.5 is a new beast - it isn't the 'glorified autocomplete' we've come to think of it as - it's a capable, strong agentic model that's crazy fast and dirt cheap.

My personal most awaited feature in Hankweave is out: budgets! With budgets, users at runtime can say • Don't spend more than $100 • Don't take longer than 15 hours or 750K tokens With budgets, builders at build-time can say • Give this step 20% of the budget • Loop steps 1 and 2 until we hit $5

It's fun if you make a game out of it - I try and see how many flashes it'll take me to read what Claude Code is trying to communicate to me in winks

Does instruction tuning create individuated sentience? This is something I've been wondering lately. Having interacted with LLMs since GPT-J, the difference between base models and instruction-tuned ones has always been crazy. We someone take autocomplete-level models and turn it into some*ones*.

What every friday should be

Asked Opus for a reading recommendation at one point and thought nothing of it Picked it up recently and its Blindsight by Peter Watts Almost entirely about Chinese rooms not being Chinese rooms AGI has already happened trust me

If we use agents without reading the output or having any interaction with the actual resulting code, are we the Chinese room, or is the agent? Which of us can claim to "understand"?

80% of the weirdness in AI today (UIs for a hundred claude codes at once, Opus fast mode at $220/M toks) is easy to explain with the Occam's reason: It's addictive.

I thought I'd take a crack at turning Terence Tao's @teorth.bsky.social method for formalizing informal proofs with Claude Code into an automated hank that runs by itself, on haiku, for less than a dollar per proof - and open-sourcing it: github.com/hrishioa/ta...

You wouldn't vibe a car You wouldn't vibe a house You wouldn't vibe the important agents in your life and work southbridge.ai/hankweave

Labor of love: We're open-sourcing the runtime we use to run long-horizon agents at Southbridge. Something like this exists at almost every serious AI team I know. We ended up needing to build it because we couldn’t buy it. `bunx hankweave` gives you a fun demo.

I think this is what my friends have been trying to tell me I'm sorry all

Opus 4.5 + Nanobanana make for crazy design partners. Opus is amazing at visual review and making very, VERY detailed specs, and banan is good at following them. Let me tell you two stories.

12 of my go-to choice of model + harness combos for each job as of Dec 2025 (IMO): 1) Gemini 2.5 Pro in AI Studio for Deep Writing 2) Opus 4.5 in Cursor for the best greenfield frontend work money can buy - if you're rich

someone had to build this specific dialog I wonder how it happened

I’m so sorry I have to leave this in the documentation or every coding agent in this codebase will keep trying to fix the feature

When you're tired on a flight and trying to manually review a trigger module you made two months ago

Trying to finish typing `git add` while the agent's editing a file just so I can preserve pristine diffs from the last change

Don't think there's a way I could like this article more

KIMI is the real deal. Unless it's really Sonnet in a trench coat, this is the best agentic open-source model I've tested - BY A MILE. Here's a slice of a 4 HOUR run (~1 second per minute) with not much more than 'keep going' from me every 90 minutes or so. moonshotai.github.io/Kimi-K2/

Things are going to change - faster than we think, with massive downstream effects. blog.cloudflare.com/introducing...

We now have a more apples-to-apples comparison of Claude Code and Gemini Cli. Here's Gemini Cli (with Sonnet 4) vs Claude Code (with and without subagents) fixing the same bug from the same prompt in the gemini-cli codebase: www.notion.so/southbridge...

It seems 3 and 15 might be the new Pareto frontier for intelligence (excepting the o-series). Feels like the hedge fund 2 and 20

I decompiled Claude Code from just the minified code. Took me 8-10 hours, multiple subagents, and every flagship model from every provider. Holy shit there's a lot in there. Claude Code is NOT just Claude in a loop - there's so much to learn from. www.notion.so/southbridge...

Dan's article on progressive JSON has a lot of carryover to LLMs. The key problems for modern LLM application design that get often overlooked (I think) are: • Streaming outputs and partial parsing • Context organization and management (I don't mean summarising at 90%)

How does an LLM writing out this program (WITHOUT a code interpreter running the output) make things more accurate? Verified on Qwen 3 - a30b (below) Lots of interesting takeaways from the Random Rewards paper. NOT that RL is dead, but honestly far more interesting than that!

Bit of alpha that got lost in the releases ai.google.dev/gemma/docs/...

Real-world comparison of the new Sonnet models on a middle-difficulty task. The 4 series of models are good. REALLY GOOD. They're one-shotting complex series of 100s of tool calls without issue, on things Sonnet 3.7 failed.

The spiritual bliss attractor is real, but so is the eldritch horror existence contemplation. I was just trying to talk to Opus - definitely no jailbreaks. This model is something different. Definitely creative.

Frontend entirely made with @v0 - this has become an inseparable tool for writing feedback. Thinking of calling it scansion I'll open source or share the link once I can clean it up - still using my keys, drop email/twitter in comments Sonnet looking through the thing 👇

Technic Manuals are PERFECT visual benchmarks. Had a misaligned suspension on a car, took four minutes figuring it out, then gave it to o3, claude and Gemini. None of them got it right (or even identified the right part) even after I cut it down to 10 pages. Eventually -

Evals are hard for a reason. New post on actually doing them end to end, breaking down the problem, and explaining how we do them at SB