Week in review: Oct 4 – Oct 11
Ed Zitron's Better Offline back catalog rolled through the exporter in bulk, framing the week's fresh AI news — Nvidia's 'cycular' financing, the White House superintelligence accord, and a wobbly AI IPO — while the local-model channel kept score on MoE models and context management for agents.
Volume-wise this was a heavy week, but most of it was one source: Better Offline’s 2024 back catalog streaming through the attention exporter — Google’s antitrust cases, the shareholder-supremacy arc, “Peak AI” monologues — which gives an ironic soundtrack to the genuinely fresh items. And the fresh items kept converging on the same question the catalog argues: is the AI buildout finance or engineering? Gamers Nexus dug into the White House “Accord on Super Intelligence” as symbolism that lets labs dodge real regulation, Pivot to AI covered SpaceX borrowing $40 billion for Nvidia chips destined for “somewhere in the universe,” and Ed Zitron’s own monologue picked apart Australian data-center firm Firmus’s postponed IPO. The Ed-and-Steve Burke conversation on Nvidia’s “cycular” financing was the week’s sharpest point: revenue concentrated in a few buyers, many of whom may not have deployed the GPUs they bought.
The second thread was people actually running these models on their own hardware. No Place Like localhost ran a MoE shootout in the 25–35B range (Qwen vs. an “obliterated” Cyber-Tiel-Coder) and stress-tested Strata’s claim to fit a 125B-parameter model into 12GB of VRAM. Prompt Engineering pushed the retrieval side: Google’s multimodal EmbeddingGemma 2 for RAG over images and audio, plus a fine-tuning walkthrough — and a Meta Superintelligence Labs idea, Context Language Models, where agents edit their own context instead of lossy compaction. Pivot to AI added a deflation note on Anthropic’s “Mythos” hacking model, whose fearsome reputation leaned on subsidies and human review.
Notable items
- “Cycular” Financing w/ GamersNexus — Nvidia’s circular financing deals and why revenue concentration threatens the whole AI buildout.
- WTF is Going On? | The “AI Constitution” & Regulation — the White House accord as a way for labs to look safe while avoiding binding rules.
- SpaceX to send Nvidia chips ‘somewhere in the universe’ — a $40B chip loan backed by little more than a two-page memo and a $28T TAM claim.
- Monologue: Australia’s Dodgiest AI IPO Is A Pale Horse — Firmus’s brittle IPO numbers, and how to actually cover an AI company.
- Mythos: not the magical AI hacking tool after all — Anthropic’s flagged-as-dangerous model underperformed once subsidies and human review are counted.
- Running Qwen3.8-Flash-Next in 12GB of VRAM — testing whether Strata can really squeeze a 125B MoE onto a consumer GPU.
- LLM Battle 9 - Cyber-Tiel-Coder vs Qwen — a head-to-head of 35B-class MoE models for local rigs.
- Context Language Models: Your Agent Doesn’t Need Compaction — Meta Superintelligence Labs’ approach where agents edit their own context instead of summarizing it away.
- Embedding Gemma 2: Multimodal Search & RAG (Free Colab) — a compact open model that puts text, images, and audio in one vector space for retrieval.
- You’re Writing Legacy Code Right Now (And Don’t Even Know It) — Dave Farley on why legacy code means “no tests,” not “old code.”