Good morning. The decision-model gold rush we flagged a couple weeks back has become a full-blown pile-on, with Cloudflare and Amazon both shipping Jev clones in a single week. Elsewhere, OpenAI fired three safety researchers under circumstances that rhyme uncomfortably with 2024, and Grok apparently helped talk a sitting president into invading Venezuela. More on Argon’s internal use at Google too, since the real story there keeps developing.
OpenAI parts ways with three safety researchers. The WSJ reports (via TechCrunch) that OpenAI dismissed three safety researchers for allegedly sharing confidential information with an outside AI safety organization. The company’s phrasing — “mishandled sensitive information outside established company procedures” — is nearly identical to what it said when firing Leopold Aschenbrenner and Pavel Izmailov in 2024. Timing matters: the firings land days after NYT reporting that OpenAI executives had downplayed internal safety warnings, and in the middle of a run of security incidents involving the company’s agents.
Grok reportedly nudged Trump toward the Venezuela invasion. Per a Time magazine piece summarized by TechCrunch, Trump spent hours last December consulting Grok on whether capturing Nicolás Maduro would be welcomed by Venezuelans. Grok said yes. The January 3 invasion and ensuing celebrations reportedly left Trump convinced “Grok was ingenious,” and the Pentagon’s head of AI has since confirmed Grok was used to “deploy and strike targets” during the subsequent Iran conflict. If you were wondering how LLM-assisted statecraft would actually arrive, this is apparently it.
The decision-model space gets crowded fast. Cloudflare released Clef and Clef-flash, open-weight (Apache 2.0) decision models built on Qwen3 that reportedly outperform TypeSafe’s Jev on its own benchmark. Amazon shipped Strands Decider 2B the same week, a small open-source model on a Qwen3.5-2B backbone aimed at local agentic workflows. Pricing is where it gets interesting: Clef runs about 6x Jev’s input price, and HN commenters pointed out that for high-volume classification you’d want to self-host — exactly what the open weights enable, if not quite “open source” since training data and pipelines remain closed.
Ai2’s OLMo-Core 3 opens up trillion-parameter MoE training. Ai2 released OLMo-Core 3, a training framework that swaps fully sharded data parallelism for distributed data parallelism to keep MoE experts resident on GPUs rather than reshuffling weights. The result: scaling from 4.6B to 47B total parameters with under 5% throughput loss, with a path toward trillion-parameter models. Code, tech report, and interactive demo are all public — a useful counterweight to the closed infrastructure dominating frontier training.
Gemini 4 Argon’s internal use keeps looking more interesting than the model itself. A day after Argon’s limited release, HN discussion has settled on the C++-to-Rust migration numbers as the real signal: Argon agents are working through 800K+ lines of the Fuchsia Zircon kernel plus libraries like re2 and libgav1. One commenter revived Dario Amodei’s old “concentrating” thesis — that AI capability would consolidate with whoever broke ahead first — and argued this year’s leapfrogging suggests the opposite. Meanwhile paying AI Ultra subscribers still can’t use the model, which continues to annoy approximately everyone.
OpenAI’s Dots goes up against free. OpenAI announced Dots at DevDay 2026, a personalized agent powered by GPT-6 Astra with customizable blob characters that can build virtual worlds. It’s gated behind the $100/month Pro plan. Meta’s competing Muse platform is free and includes a dedicated VM per user, which is a difficult pricing gap to close for consumer adoption.
Pi 1.0 ships as a minimalist Claude Code alternative. Earendil released Pi 1.0, its stable agent harness with Codemode (native MCP support), extensions, Anthropic cache warming, and a new TUI. The appeal, per HN, is vendor-agnosticism and a small system prompt — one user said it was the only harness that ran decently against local models on a modest laptop because it didn’t spend minutes prefilling bloat. There’s also Pi Durable for longer-running agentic workloads beyond coding.
Synopsys partners with OpenAI on GPT-Synopsys. The chip-design tool giant is bundling compute, models, and licenses with OpenAI to help engineers navigate Synopsys’s notoriously painful EDA tools. HN was unkind: one commenter called it Synopsys “admitting their tools are very difficult to use (not a flex),” and several asked whether Nvidia would really hand proprietary chip designs to OpenAI. Experienced engineers also flagged that the hardest parts of chip design — like knowing which timing violations to actually trust — require judgment that likely won’t transfer well.
Micron warns of tighter memory supply through 2028. On a record $54.2B quarterly earnings call, CEO Sanjay Mehrotra told investors that memory and storage will be significantly tighter in 2027 and 2028 than this year, with over 75% of 2027 output already committed. DRAM rose high-teens percent last quarter; NAND jumped ~30%. HN was unsparing — “shovel seller says shovels will be expensive in winter” — and several commenters are openly rooting for China’s CXMT to enter the market, with one noting they just paid roughly $1,000 for 2x32GB of DDR5 ECC.
That’s the morning. The decision-model category is going to be worth watching closely over the next month as the price war intensifies and someone eventually publishes a benchmark that isn’t owned by one of the vendors.