---
title: 'In the News: November 20, 2026 (Evening)'
description: 'Ankit Gupta let coding agents run for days on an AWS FPGA and got a language model 33 times faster, held to a bit-exact reference.'
canonical_url: 'https://darkfactory.dev/news/2026-11-20-evening'
markdown_url: 'https://darkfactory.dev/news/2026-11-20-evening.md'
collection: news
date_published: '2026-11-20T19:06:00-05:00'
date_modified: '2026-11-20T19:06:00-05:00'
---

# In the News: November 20, 2026 (Evening)


Ankit Gupta pointed coding agents at an FPGA, a kind of hardware he says he had never programmed, and left them running for several days. They got a language model working on it and made it 33 times faster. What kept that unattended loop honest was one rule: every result had to match a reference implementation bit for bit.

## 1. Coding agents spend days and $10,000 getting a language model 33x faster on an FPGA

**[Vibecoding an inference chip](https://ankitg.me/blog/vibecoding-inference-chip/)** · Ankit Gupta · Personal blog, November 20, 2026

Gupta rented an AWS F2 instance, which has an FPGA with 16 GB of HBM, and set out to run DeepGrove's ternary-weight model on it. A cloud agent ran in Conductor for several days while he mostly encouraged it. He says he has never written Verilog. By his account the loop ran hundreds of experiments, wrote roughly 3M lines across its history, parsed 30M lines of output, and cost $10,000 in Claude and OpenAI tokens. Throughput went from 6.43 tokens per second at the first full-model bit-exact baseline to 213.13 after 14 steps. Most steps added a modest gain rather than one big jump.

His November 15 comparison table puts the final FPGA image at 213.13 tokens per second on an f2.6xlarge at $1.98 an hour, matching the pinned CPU reference on 24 of 24 turns. The fastest exact CPU setup in the table, 32 threads on an m8a.8xlarge at $1.95 an hour, managed 129.5. A community CUDA port on a T4 reached 166.3 but matched on 0 of 24 turns.

The post includes the agent's own retrospective, written when Gupta asked which pasted design advice it had actually used. It rejected all five numerics shortcuts in that advice because each changed the output bits, and "your bit-exact rule ruled it out." The advice to stop running jobs that take hours had a quicker effect: the agent dropped a simulation that had run for about 24 hours. It adds a cost it says it cannot prove: two hangs that only RTL simulation would have shown reached the board, and the integration lost 38 hours to them. Gupta's conclusion: "the humans are still useful at defining how to iterate fast and what correctness looks like, and the agents are great to manage the exploration." He says he would not recommend this setup for production. The post links no code, so the figures are his own and we have not seen them reproduced.

**Why it matters:** A bit-exact check against a pinned reference let this loop run for days with little steering, and it also decided which design shortcuts were allowed. Anyone planning a long optimization run for an agent should write that check before the first prompt. The 38 hours is the agent's own unproven estimate, but it is a reason to decide in advance which slow checks a loop may drop.

---

## Also this cycle

- **[OpenComputer](https://github.com/diggerhq/opencomputer)** · diggerhq, GitHub · An Apache-2.0, self-hostable coding agent that you assign work by mentioning it in a Slack thread. It runs Claude Code, Codex or OpenCode in sessions isolated with rootless Podman on a single VM, and the result arrives as a pull request. The team's own benchmark of 4 PRs says it opened them with CI green in 56 seconds against 185 for Devin.
