Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

AI Trending Briefing · November 8, 2026 · afternoon edition

This weekend's agent tools make deliberate choices about what an agent holds, with Google's Elephant Goldfish Model keeping reviewers out of the conversation, billion-context folding old turns into summaries, and llama-modchip squeezing a 284B model and its cache into 64 GB of RAM.

llama-modchipElephant Goldfish Modelbillion-contextGitHub CopilotGoldman Sachslocal-aiagent-memoryai-skillsagent-harnessclaude-code

Trending AI Briefing: Sunday, November 8, 2026 (afternoon ET)

Sunday is quiet on the vendor blogs and loud on GitHub. The projects climbing the boards share one habit: they make deliberate choices about what an agent holds. Google's Elephant Goldfish Model reviews a design with agents that are deliberately kept in the dark about the conversation, billion-context squeezes old turns into summaries so one session can run for months, and llama-modchip packs a 284-billion-parameter model and its cache into a 64 GB desktop with no GPU. The other half of the day is about who gets an agent at all.

What's hottest in AI news right now

llama-modchip says it runs DeepSeek-V4-Flash, a 284B mixture-of-experts model, on a 64 GB PC with no graphics card. The project reached Hacker News at 10:22 UTC on November 8. It is a set of patches to llama.cpp plus its own quantizer: the README on its main branch lists a 56.56 GiB build, roughly 58 to 59 GiB of memory in use at full 256K context, about 44 tokens per second reading a prompt, and about 9 to 11 tokens per second writing on CPU alone (up to about 14.7 with its built-in draft-and-verify speedup), measured on the author's own 16-core Ryzen 9 9955HX. The latest tagged release, v2.4.0, is dated November 4, and main has moved on since. The catch is in a compatibility note: its custom tensor type (ID 40) makes unpatched llama.cpp, Ollama and LM Studio fail with unknown tensor type 40, so the files only run in modchip's own binaries. Its ./modchip claude launcher also starts Claude Code with --dangerously-skip-permissions, which turns off approval prompts. Every speed figure is self-reported or community-submitted. README · v2.4.0 release · HN

Google's Elephant Goldfish Model (EGM) shipped v0.1.1 on November 6, its second tagged release after v0.1.0 on October 27. EGM is a plugin for Claude Code, Codex CLI and Antigravity CLI that splits design work into roles: the "Elephant" agent holds the project history and drafts a spec, and "Goldfish" reviewers are fresh subagents that see only the draft and its sources, hunting for unstated assumptions. Seven skills cover brainstorming, drafting, review, a review loop for non-software docs and a drift report against current code. The README calls it "not an officially supported Google product" and says plainly that "A review does not replace tests or human judgment." Shields shows about 15k stars. Repo · Releases

@GitHub in Slack and Microsoft Teams is generally available as of November 6. From a chat thread you can investigate a bug, open an issue, have the agent implement a change and get a pull request back. The release adds "Canvas in Teams" for reviewing plans and code beside the conversation, and the agent now picks the repository from your message instead of asking. One small detail matters more than it looks: issues and PRs it creates now carry the line "Created from a Slack/Teams conversation with @username," which gives reviewers a trail back to the person who asked. GitHub changelog

A report that Goldman Sachs pulled Claude Code back to its engineers reached the Hacker News front page today. The story is not new. State of AI's November 2 write-up relays Bloomberg News: internal guidance updated in October "restricts Anthropic's Claude Code and Microsoft's GitHub to the bank's engineers," after "a review of productivity and risk concerns," cutting off traders and bankers who used the tools to track client flows and build market models. Strats keep access, per a second unnamed person. A Goldman spokesperson said only that the bank makes certain AI tools available "to determine the best use cases." The Bloomberg original did not load for us, and State of AI is published by Outlever, a brand-newsroom company, so treat the reasons beyond "productivity and risk" as that outlet's inference. State of AI · HN

OpenAI raised about $11 billion at a $1.25 trillion pre-money valuation, per 24/7 Wall St. The November 7 report attributes the $11 billion, which reached OpenAI's accounts by Friday November 6, to "people familiar with the deal," and says CFO Sarah Friar confirmed the $1.25 trillion figure in a TV interview, below the $1.4 trillion in earlier press reports. SoftBank and MGX took part, a separate report puts a further tranche of up to $9 billion in play, and the round could reach $20 billion. OpenAI declined to comment on the funding. 24/7 Wall St

New tools and features worth actually trying

Elephant Goldfish Model in Claude Code. claude plugin marketplace add google/elephant-goldfish-model, then claude plugin install egm@egm, start a new session and run /egm-draft on your next spec. It gives you a reviewer that cannot be talked into agreeing by forty turns of shared history. Honest tradeoff: it needs a harness that can spawn fresh subagents (EGM stops the review if it can't), and the built-in workflow targets software specs only.

llama-modchip's hardware check before any download. The README offers check_cpu_readiness.py to tell you whether your CPU, HugePages and shared memory are ready, without pulling a 56 GiB model. Honest tradeoff: the documented command pipes a script from the internet straight into python3, so download it and read it first, remember the models it leads to will not open in Ollama or LM Studio, and know that its Claude Code launcher skips permission prompts.

vercel-labs/agent-browser. npm install -g agent-browser, then agent-browser install to fetch Chrome for Testing, and your agent drives a browser from the shell using short text snapshots with element references like @e1. agent-browser mcp exposes the same thing as an MCP server. Honest tradeoff: it is labeled a Labs product, and element references go stale after any navigation, so agents must re-snapshot.

GitHub's new pull request Overview page. An opt-in public preview from November 5 puts merge status and next actions at the top and filters comments by All, Unresolved or Mine. Click "Try the Overview page" in a PR header. Honest tradeoff: rollout was still continuing through the weekend, so not every account has it, and it changes nothing about who must approve. Changelog

Trending AI repos on GitHub today

Read from Trendshift's daily board at about 15:10 ET on November 8; its figures are momentum scores, not star totals. Stars below come from cache-busted shields.io badges, which round.

  • google/elephant-goldfish-model (#16): draft-then-review design skills where reviewers start with no conversation history. Why now: v0.1.1 landed this week. Apache-2.0 (the LICENSE file has no copyright line filled in), about 15k stars, v0.1.1 November 6, not an official Google product.
  • vercel-labs/agent-browser (#12): a Rust CLI that lets agents drive Chrome through shell commands and compact page snapshots. Why now: steady releases and an MCP mode. Apache-2.0, about 47k stars, v0.41.2 November 3, badged as a Vercel Labs product.
  • ccch1mneyyy/dsh-Go (#10): an unofficial Android client that runs DeepSeek Harness on the phone or remote-controls it on a computer. Why now: frequent releases. AGPL-3.0 (sdk and examples MIT), about 16k stars, v1.0.13 November 7. You bring and pay for your own model provider, anonymous usage stats are on by default, and phone-control mode can send screen content to the model.
  • ranxianglei/billion-context (#25): a proxy between a coding agent and its model API that folds older conversation into summaries. Why now: it claims a 100K window is enough for month-long sessions. MIT plus an extra attribution term for products with user-facing surfaces, about 5.1k stars, v0.1.213 November 7, default branch master. All benchmarks are the authors' own, the README's status section still calls it "Early," and auto-update checks npm every three minutes unless you set ACP_AUTO_UPDATE=0.
  • storytold/craft-anything (#24): a Rust engine that studies an existing app or site and has a coding agent build and check a new one. Why now: v0.5.0 shipped November 6. MIT OR Apache-2.0, about 22k stars. Needs a local coding agent, and images or 3D need your own fal.ai key.
  • anthropics/skills (#23): Anthropic's example Agent Skills. Why now: back on the board as skills spread across harnesses. About 189k stars, no releases, and no root LICENSE file; the README says many skills are Apache 2.0 and that they are for demonstration.
  • synchopate/llama-modchip (Hacker News, not on Trendshift): CPU-only inference patches for large mixture-of-experts models. Why now: today's local-AI thread. Apache-2.0, about 400 stars, v2.4.0. Benchmarks are self-run, and its custom tensor format only loads in its own binaries.

What actually matters from today's signal

The trend to track is that deciding what an agent holds is becoming a design choice with its own tools. EGM bets that forgetting on purpose makes a better reviewer. billion-context bets that compressing on purpose keeps one session alive for months. modchip bets that aggressive quantization and cache-aware kernels make a 284B model fit in desktop RAM. For builders, the high-signal areas are reviewer isolation for specs and plans, compression layers you can inspect and turn off, local inference that runs on memory rather than graphics cards, and agent output with provenance attached, like GitHub's "Created from a conversation" line.

The counter-signal is that every one of these bets ships with the maker's own numbers and a dependency you might not notice. modchip's speed lives in a file format only modchip reads. billion-context sits on your model traffic and updates itself. EGM's quality claim has no published evaluation. Meanwhile the access question is getting real: GitHub is putting agents in the chat rooms where non-engineers work, and the Goldman report points the other way, toward agents for the people who can read what the agent wrote. The teams that get this right will decide who reviews agent output before they decide who gets an agent.

Source access notes: Vendor pages, Hacker News (Algolia, stories since about 08:10 UTC November 7, plus the front page), Trendshift and repo files were read from about 15:10 to 15:40 ET through fetch agents, because neither the workspace shell nor the Mac shell could reach them. Anthropic, xAI and the Microsoft Agent Framework blog had nothing new in the window; blog.google served a stale 2025 listing; DeepMind showed month-only dates, so Gemini Omni 1.2 Flash is left out. openai.com/news/ failed and the RSS fallback was used; the November 6 Hacktron partnership page did not load, so it is left out. The fa-mag copy of the Bloomberg Goldman story did not load. The Codex changelog redirects to another domain and was not followed. GitHub releases.atom feeds were rate-limited for several repos, so some release dates come from releases pages that show no year. A Sonnet 5.5 adversarial pass checked the draft and caught ten issues, all fixed: EGM v0.1.1 is its second tagged release (v0.1.0 was October 27), an unsupported Hacker News point count, a bits-per-weight figure the README does not state, a misquoted modchip compatibility note, a memory figure that kept only the low end of the README range, a speed range that blended CPU-only and draft-assisted decoding, README numbers coming from main rather than the v2.4.0 tag, OpenAI sourcing (Friar confirmed the valuation on TV), an overstated "sponsored" label on State of AI, and a thesis that overstated how far the three projects converge. While drafting the article on llama-modchip, a Codex fact check found that its ./modchip claude launcher passes --dangerously-skip-permissions; we confirmed it in the script on main and added it here. Claude Code's npm latest is still 2.1.319 (00:58 UTC November 6). Stories covered in the November 7 and November 8 morning briefings (Apiiro's .claude secrets count, MaxxBench, LoreKit, "Goodbye Claude, Hello Pi", Anthropic's research-use study, Pinky, Microsoft Research AI Frontiers, LangChain async subagents, Cloudflare agent identity) are not repeated.