curated · 2026-06-29

Two numbers make this worth clicking: 1.6 trillion parameters, trained entirely on Chinese ASICs, and it's not a lab flex—it's the #1 model on OpenRouter by actual usage, with users burning through 10 trillion tokens. Owl Alpha ranks #1 on Hermes Agent, #2 on Claude Code, and #3 on OpenClaw, so people are genuinely using it for agent and coding tasks. I'd discount the performance claim a bit. "Gemini or Opus 4.6 level" comes from the official tweet; the post doesn't share specific benchmarks or scores, so there's no way to verify. The ASIC details—chip model, interconnect, training stability—are also missing. And the model is retiring soon with no timeline for the next version, so jumping in now means a likely migration. If those rankings hold up, the real signal is this: Chinese ASICs can now train a commercially viable giant MoE that people actually pay to call. That's harder evidence than any benchmark number.

The timeline is what makes this worth clicking: just three months after GTC 2026, NVIDIA has scrapped the 4-die Rubin Ultra package and replaced it with a half-size, half-performance version. SemiAnalysis pins it on manufacturing execution problems, but the post doesn't say whether it's TSMC's CoWoS or another packaging step that failed, and there's no new delivery timeline. I'd discount this a bit since it's a single tweet with thin details. But if true, the impact is real: the original Rubin Ultra was the biggest compute upgrade after Blackwell, and halving its performance means everyone's 2027 training cluster math needs a redo.

This one's worth opening because Piotr Migdał's test is concrete: 8-bit quantized Qwen 3.6 27B on a Macbook Max M5 128GB, running via llama.cpp with multi-token prediction, hitting 32 tok/s and using 42GB RAM. He had it generate a hexagonal minesweeper npm package from a single prompt, plus a poem mixing quantum physics and Zouk dance. I'd discount this a bit — it's a personal blog experience, not a systematic eval. But 32 tok/s and 42GB RAM on a laptop is genuinely usable. He recommends llama.cpp over Ollama on ethical grounds; the post links to his reasoning, so you can judge that yourself. If you've got a 128GB Macbook, the post gives you the full launch command to reproduce it. What's missing: head-to-head comparisons with other models in this size class, and how it holds up at longer context lengths.

The LA Times investigation breaks this into two layers. The bottom layer: MIDB, a targeting database built in the 1980s, was never connected to the annotation tools analysts actually use. An analyst flagged the building as a school in 2019, but the note stayed in a separate system. Commanders reviewed the site using seven-year-old imagery. The top layer is a speed mismatch. Anthropic's Claude, embedded in Palantir's Maven system, suggested roughly 1,000 targets on day one. Human vetting capacity was already hollowed out — Pentagon AI lead Jack Shanahan said he couldn't fill targeting roles as early as 2017. Don't read this as 'AI picked the wrong target.' The more accurate picture: AI accelerated throughput in a pipeline where the final human check was never scaled up. The LA Times mentions hopes for automated cross-checks against Google Maps, but the article doesn't say if that has a timeline. What's confirmed: the MIDB-to-MARS replacement is years behind schedule.

This is worth a look because Cursor connected cloud agents to a phone in a way that fits real workflows. You can kick off an agent with voice or text while you're at lunch, let it investigate a bug or tweak a UI, then review diffs and merge PRs right in the app. The team already uses it for on-call incidents — that's more practical than the old "code on your phone" pitch. Don't read this as "phone replaces laptop." It's a remote control: cloud agents run in isolated VMs with full dev environments and can hand work back to your local machine. The real win is firing off tasks in spare moments and reviewing results later, not building projects from scratch on a screen. The post doesn't disclose pricing. It says Composer 2.5 runs are 75% off through July 5, but the base price isn't stated. I'd treat this as a mobile control panel for paid users — free tier doesn't get access.

The useful bit here is the granularity: RedKnot pushes KV Cache optimization from the layer level down to individual attention heads. It found that 83.4% to 96.8% of heads only care about local context, so it stops storing full global KV Cache for every head. Instead, it applies head-specific sparsity, pairs it with sparse FFN and SegPagedAttention, and aligns the algorithm with the storage pattern. On 8×H800, TTFT speeds up by up to 3.54×, per-card concurrency jumps 7.8×, and prefill FLOPs drop 67%–79.5%. On DeepSeek-V4-Flash with 128K context, TTFT is 5.16× faster and KV transfer shrinks up to 6.3×, while accuracy stays above 95% of dense F1. I only have the summary—no task details or eval sets yet—but the approach looks practical, especially for long-context serving where KV Cache eats VRAM fast.

The numbers here are worth sitting with: $590B across four fabs, a packaging center, and next-gen R&D, from two companies that already control ~80% of the HBM market. Jefferies' price forecast is specific — 40–50% in Q3 2026, another 30–40% in Q4, then 40–45% in 2027, with no relief until 15–20% of new capacity comes online in 2028. Apple hiking Mac prices tells you the cost pressure is already moving from data centers to consumer hardware. I'd discount the headline a bit: the $590B is spread over 15 years, not a single-year spend, and the price squeeze won't ease until those fabs actually ship chips.

I clicked because the author shares two copy-pasteable prompts, not just hand-waving about first principles. The first forces the AI to rebuild reasoning from base facts instead of analogies—it once surfaced a hidden traffic-routing flaw that triggered a full rewrite. The second makes the AI attack the code as a malicious user, catching OOM infinite loops and future-timestamp pollution that manual review would likely miss. Together they create a generate-then-verify loop; the project handled 10M+ requests last week. I'd save both prompts and try them immediately, but note: the post only gives paraphrased descriptions, not the full prompt text, and results probably depend heavily on your task type.

Tunguz lays out a cost structure that's hard to ignore: Anthropic burns $2M per employee on compute, while the top 1% of software companies spend $89k per engineer and the median spends $137 a year. The gap is wild. He maps three 2029 scenarios. Bear case: token prices keep falling 10x/year and open-weight models close the quality gap, so AI spend settles at 41% of engineer salary. Bull case: agentic workflows drive a 24x jump in token consumption, and the whole market converges to Anthropic's current 230% ratio — $596k per engineer per year. I'd discount the precision a bit. All three paths lean on Ramp's spending index and Goldman's token forecast, but the post doesn't say which one to model for 2027. The Bull case also assumes frontier model prices hold steady and supply stays tight, which already feels shaky in mid-2026. The real question isn't the exact number. It's what happens when the AI bill starts looking like another headcount. Do you ration tokens per role, or do you hire fewer people and run more agents? That tradeoff is the thing to watch.