curated · 2026-06-07

Nvidia’s multi-year pact with SK Hynix is less about a named memory part and more about pulling the memory bottleneck inside Nvidia’s planning loop. The article gives only co-development across future AI memory generations; specs, production timing, and financial terms are absent. So I would not treat this as a confirmed HBM4 or custom-stack launch yet. SK Hynix already sits closest to Nvidia’s HBM cadence, while Samsung and Micron are fighting for roadmap relevance, not just bandwidth bragging rights. For model labs, memory capacity, bandwidth, and allocation priority keep setting the practical ceiling on training runs and inference margins. Big headline, thin disclosure, clear direction: AI accelerator competition is shifting another notch toward memory control.

OpenAI’s redesign smells like margin pressure: 900M weekly users, 50M paid users, $2B monthly revenue, and still no profit. The chat box no longer carries the growth story. Pulling Codex, image generation, Canva, and Booking into one surface turns ChatGPT from a Q&A destination into a task router. The strongest hook is Codex desktop passing 5M weekly active users. Coding remains the agent use case with cleaner pricing and cleaner ROI than travel booking or slide-making. I don’t buy the executive line that “chat is dead”; users still want chat, they just won’t keep paying premium prices for one-off answers. Anthropic has kept Claude close to developer and enterprise workflows. OpenAI is dragging its consumer funnel in the same direction now.

Tomiyasu’s case lands because it is not an “AI for agriculture” fantasy. It is one farmer using Codex as glue across 8 scrappy workflows: disease photos, NDVI checks, ESP32 curtain control, LINE bots, sowing counts, RTK-GPS cost study, and an Airtable database. I don’t buy the “super engineer” framing literally. Codex did not bring farm judgment, sensors, machinery, or local know-how. The sharp part is that the integration tax collapsed. Work that once needed a small SI vendor now gets hacked together by the domain expert who owns the problem. Vertical SaaS vendors should hate this pattern: the user bypasses sales first, then implementation.

Harness-1 is a shot at hand-built retrieval pipelines, not another query-rewrite trick. UIUC and Chroma trained a 20B subagent on gpt-oss-20b with reinforcement learning inside a stateful search harness. They report 0.730 average curated recall across 8 benchmarks, 11.4 points above the next open-source subagent, and behind only Opus-4.6. If that reproduces, a lot of bespoke RAG routing code starts looking brittle. I have one concern: the article gives the headline benchmark, not the ugly deployment details. Pricing, latency, private-corpus behavior, permission filtering, and failure cases are not given. That is where retrieval agents usually bleed.

Application teams are learning to cut model tax with routing and SFT, and frontier closed models are losing default status. Harvey moved Kimi 2.6 from 11% to 15% all-pass on a 100-task Legal Agent Benchmark slice, beating Opus at 14%. The cost gap was the punchline: $84 versus $954 for the same 100 tasks, roughly 11x cheaper. That is sharper than the usual “open source is good enough” line. It says vertical workflows can use post-training plus their own eval loop to drag cheaper models into the replacement zone. Coinbase says costs stayed roughly flat while token use grew exponentially. Lindy says it moved 100% of traffic to DeepSeek v4. I’d still haircut the Harvey claim: the article gives a 100-task slice, but not LAB contamination controls or fit to live legal workflows.