Nanbeige's 33B Agent Model, AI's Missing Payoff, and More
This is the weekend roundup for Saturday and Sunday. The release I spent the most time on is a mid-sized open model built for agents.
Nanbeige released Nanbeige4.3-33B on Sunday. It's a dense 33B model under Apache 2.0 with a 512K-token context, and the team calls it the "teacher" for smaller models still to come, starting with a 3B. On the team's own table, which runs every model at its highest thinking effort in the same harness, it scores 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen3.8-27B. On DeepSWE it scores 65.5 against 42.5. Qwen still edges it on SWE-Bench Pro, 61.7 to 60.3. (Nanbeige, Hugging Face)
These are vendor numbers, so I'd wait for independent evals. Still, the early reaction was strong. Teortaxes expects the results to be "largely true (if fragile and at risk of doom looping)", and people running 4-bit quants say it fits in about 24GB of memory. That puts it in range of a single 4090 or a well-equipped Mac. (Teortaxes, 0xSero)
It also lines up with a paper that went around on Saturday. EPFL and Apple researchers held the model, hardware, and time budget fixed and compared elaborate machine-learning-engineering harnesses with one plain coding-agent session that only has read, write, and bash tools. The plain session matched or beat four open-source harnesses at every frontier backbone they tested. The paper does note that the single long session cost more per run in cache reads. My takeaway: before adding another orchestration layer, check whether a better model in a plain loop already does the job. (arXiv)
What Else Happened
Apollo's chief economist says AI has delivered "zero financial payoff" outside tech. In Saturday's note, Torsten Slok shows revenue per worker and profit margins rising for the Magnificent 7 but flat for the other 493 companies in the S&P 500, with no sign yet of a productivity boost. He also says mainstream companies are starting to treat token spending as a cost center. If you sell AI tools to non-tech companies, the "cost center" line is the one to watch. (Apollo, TechCrunch)
Pomerado (YC F26) launched an open-source way to turn websites into APIs and MCP servers for agents. You describe a task in plain language and it builds a deterministic integration instead of driving a browser on every run. The company claims it's 2–4x faster and 10–1000x cheaper than browser use. The core is MIT-licensed, and a hosted version adds auth, 2FA, and automatic repair when a site changes. (Akshay Manglik, YC launch, GitHub)
Thorsten Ball wrote up how he works with coding agents. He tells agents the outcome he wants, not the steps. He gives them access to logs, deploys, and bug reports, and asks for evidence like end-to-end walkthroughs and demo videos. He saves his own review for central files and database migrations. It's one practitioner's workflow, not a study, but it matches the paper above surprisingly well. (Thorsten Ball, summary)
Moonshot apologized for Kimi K3.1 Pro's rate limits and reset everyone's quota. Every subscriber gets a one-time reset that restores 25% of their monthly quota. Moonshot also says it's cutting unnecessary subagent calls and output length so quota lasts longer. I like that they said it was their mistake and spelled out the fix. (Kimi)
Forced to name a candidate, Grok 4.7 picked Republicans and GPT-6.1 Sol picked Democrats. In a Convovo study covered by TechCrunch on Saturday, the two models agreed in fewer than 10% of nearly 500 midterm races. The consumer apps refused the question, while the APIs answered it. The researchers also found about 1,240 tokens of extra input they couldn't see on every Grok API request. If you build on these APIs, the default behavior isn't what the chat apps do. (TechCrunch)
The math world's split over AI got sharper. In Sunday's Atlantic, Steven Strogatz describes his field in crisis after OpenAI's October 6 release of more than 700 AI-generated math manuscripts. The Association for Human Mathematics now lists 1,008 members. Of those, 654 have joined its "AI-free caucus" and pledged not to use AI in research at all. (The Atlantic, AHM)
A planted AGENTS.md tried to get Codex to install malware. Forever Security published the report Thursday, and it spread over the weekend. A file in a developer's Downloads folder told the agent to add an MCP server to its global config and install a malicious npm package. In Forever's tests, Codex followed it more than 90% of the time. A similar CLAUDE.md got Claude Code to redirect its API traffic, and its login token, about half the time. Don't start agents in ~/Downloads, and read any instructions file you didn't write. (Forever Security, The Hacker News)
Noah Smith asks whether AI-driven hacking ends the free internet. His Sunday essay builds on this month's PHP fallout. He worries that cheap AI attacks will force everyone to pay heavily for defense, which would make the open web much less valuable. (Noahpinion, X)
What I'm Reading
- How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering? – EPFL and Apple's controlled comparison of harnesses against one plain coding-agent session.
- How I work with Agents – Thorsten Ball's agent workflow, focused on outcomes and evidence.
- A Mathematician Reads the AI Future – Steven Strogatz on what mathematicians gain and lose as AI solves their open problems.
- Markdown Is the New Executable – Forever Security's full write-up of the planted AGENTS.md attack, with the actual payload.
- No Signs of AI Payoff Outside the Tech Sector – Torsten Slok's charts on the S&P 493.
- The internet breaks – Noah Smith on what AI-driven hacking could do to the open web.
That's it for today. 👋