Home pageBengaluruRSS

Bhavuk Manocha

Notes on AI agents, local models and making them work at work

The best reason to run a 125B model at home isn't price

Strata puts Alibaba's Qwen3.8-Flash-Next on a gaming PC, and on Alibaba's own numbers it beats February's Opus. The case for running it is control, and the narrow jobs where control matters. Cost does not make it.

94 tok/s

writing speed for a 125B model on a 12 GB RTX 5070 with Strata's smallest pack, as the project measured it

Your favorite leaderboard is benchmarking a VM

Launch-day charts make a 3-point lead look like a verdict. The labs' own footnotes show that the harness, the machine and plain cheating move scores by more than that.

8.7 pts

spread on DeepSWE for identical DeepSeek V4.1-Flash weights across eight agent harnesses (calculated: 74.2 minus 65.5)

Nobody reads the diff anymore. Did they ever?

AI agents wrote the code faster than people can read it, and the review queue is where the speed goes to die. The data says line-by-line review is fading. It also says that was never the part worth keeping.

12x

more code changes merged per day at Anthropic in four months (calculated from 2,500 to 30,000+)

Your context graph doesn't know who's asking

A graph built once, in one place, from everything the company knows goes wrong in two ways. It was built with someone's permissions, and it was built at a moment in time. The vendors' own docs list the seams.

77%

of deleted facts a GraphRAG index still returned after an incremental update (an independent bug report, October 2026)

The tractor took 41% of us off the farm. Agents start with the bottom rung.

Work has been handed to machines before, and on average it went well. Agents are now doing it to knowledge work, and robots are slowly reaching fields and factories. The pattern holds, with three differences that decide whether it ends well.

41% → 1.9%

share of the US workforce on farms, 1900 to 2000 (USDA)

How Claude Code won the agent race, and why the lead is cracking

Claude Code, Codex and Gemini CLI launched within four months of each other. Twenty months later one of them is the default. The reason is less about focus on coding than about one loop that Anthropic closed early.

39%

of professional developers used Claude Code at work in May to July 2026, up from 18% in January (JetBrains)

Build your own Instinct, or Muse, with Hermes Agent

Meta's Muse and the invite-only Instinct sell the same idea: an assistant that remembers you, lives in your chats and messages you first. Hermes Agent, an open-source agent from Nous Research, has every piece you need to build your own.

20+

chat apps one Hermes gateway can talk through, from WhatsApp to iMessage

Why plugging Claude straight into your data doesn't work

Connect a model to the warehouse, ask a question, get a confident number. Anthropic's own data team says that setup creates "a false sense of precision". Their fix is mostly not about the model.

21%

the most Claude got right on Anthropic's own analytics evals, before skills

A 35B model on an 8 GB graphics card

Qwen3.6-35B-A3B has 36 billion parameters and still runs on a mid-range GPU at around 40 tokens a second. The trick is knowing which 7% of the model has to be on the card, and parking the rest in system RAM.

92.9%

of Qwen3.6-35B's text weights are experts that can live in system RAM