all posts

The reason this catches attention: it moves Nvidia's compliance risk from corporate policy wrangling to individual criminal liability. Over the past two years, most DOJ chip-smuggling cases targeted middlemen or shell companies—direct charges against an original manufacturer's employee are rare. But the FT paywall blocks everything beyond the headline and a one-line summary. Chip model, smuggling route, and dollar amounts are all missing. I'd discount this a bit until we see the full article. If the chips were H100s or B200s, this is serious. If they were downgraded compliance versions already circulating in China, it's more likely an employee bypassing internal sales controls. Hold for the unlocked text.

The reason to click: it makes "always-on agent" lightweight—under 10K lines of Bash, no cron, no fixed checklist. The agent decides what to think about, who to reply to, and when. Laude and MIT ran Audel publicly for weeks; it starts its own projects, pings people unprompted, and even corrects its own messages. I'd discount this a bit. It's alpha research software. The team says don't share secrets, run it in a sandbox, and cap your API spend—because it thinks 24/7 and burns tokens nonstop. Multi-user shared thought stream has sharp edges too: Audel can't keep secrets, and they haven't tested what happens with conflicting instructions. The useful signal: if you want an agent that ruminates on its own and acts like a multi-threaded team participant, this framework lowers the barrier a lot. Just don't treat it as production-ready. It's more of an experimental rig for watching what persistent thinking produces.

This is worth a click because Andrew Ng's pivot is itself a signal: teaching people to use models is a much bigger market than teaching them to train models. His team analyzed over 10,000 job postings, ran dozens of interviews and surveys, and landed on four capabilities: building and deploying AI apps, software engineering fundamentals, using coding agents, and product sense. I'd discount this a bit since we only have the framework — no syllabus, launch dates, or pricing. But the framework is solid. The first skill stresses disciplined evals and error analysis loops so system behavior isn't left to chance, which maps closely to traditional MLE workflows. The second calls out vibe coders directly: if you don't understand tradeoffs, you'll get poor results from your coding agent. I buy that — LLMs reward expertise, and the gap between skilled devs and beginners has only widened. The third covers when to intervene versus leave an agent alone, plus routines for trying new tools as the space evolves. The fourth is product sense: when to ship an MVP fast and when to slow down. Overall, this reads more like a skills map for AI engineers than an ad for new courses. If the eventual curriculum goes deep on eval loops and agent collaboration, it'll be more useful than most prompt-engineering tutorials out there. For now, treat it as an industry signal, not a product announcement.

This one's worth opening because it puts a number on something people have been feeling: AI development is decoupling from English. Plicara crawled 1.87 million agent skill files and found the non-English share rose 3.3 percentage points in a single quarter—a pace GitHub documentation took a decade to match. Chinese skills sit at 6.2%, nearly double the Chinese share of GitHub docs, which suggests Chinese developers aren't just using AI, they're teaching it how to work. The wild part is how much published numbers disagree. English shares range from 65% to 99.7% across different studies. Plicara's explanation is refreshingly straightforward: nobody's wrong, they just sampled different populations. A healthcare skill set, a curated marketplace, and an English-seeded crawl will naturally produce different language distributions. They even ran their language detector against a competing tool and found only 1.2 points of disagreement—far smaller than the 7+ point gaps driven by data source choices. I'd hold one question open: the post doesn't touch whether non-English instructions degrade agent performance. This is a file-crawl study, so it can't answer that, but it's the thing that matters most for anyone building on this. If Chinese skills produce lower task success rates, that 6.2% number reads differently.

The timeline here is tight: Leopold Aschenbrenner left OpenAI, raised money loudly, named the fund Situational Awareness, then nearly got wiped out in the late-July AI selloff. Now the SEC is asking banks to preserve trading and funding records. No charges yet, and the fund says scrutiny is expected. I'd discount the drama a bit—SEC subpoenas after a near-blowup aren't unusual—but the investigation likely focuses on leverage, redemption pressure, and what investors were told. The article doesn't give fund size or loss figures, so we're missing the scale.

This one's worth opening because it answers "do discounts actually work" with real routing data. OpenAI ran big discounts on Terra and Luna from late July to mid-August, and the effect was immediate: Luna daily tokens jumped 13.8x, Terra 5.6x, while the undiscounted Sol barely moved at 1.1x. The punchline is where the share came from — roughly three-quarters was taken from other labs, pushing the OpenAI family's token share on OpenRouter from 7.1% to 12.4%. Retention looks decent too: of the 100K+ users who tried Terra or Luna, 32% kept using them after discounts ended, and 18% ran at or above their discount-period pace. But the post-discount window is only 6 days vs a 19-day program, so that number could shift. I'd read this as a clean case study in price elasticity for API models, but don't extrapolate everywhere. OpenRouter users are price-sensitive with low switching costs — not the same dynamic as enterprise contracts.

This piece is worth opening because it connects three long-context solutions from different layers of the stack, all driven by the same intuition. YaRN partitions RoPE frequencies—high frequencies untouched, low frequencies compressed—letting Llama 2 7B hit 128K context with ~384 A100 GPU hours. DeepSeek V4 uses full-attention within a 128-token window and heavy compression plus sparse selection beyond it, cutting V4-Pro's per-token FLOPs to 27% of V3.2 at 1M context. DeepSeek-OCR compresses full pages into dense visual tokens, hitting 97% text accuracy at 10x compression. Different teams, same trade-off: spend compute on what's close, compress what's far. The post also flags a new problem—maintaining positional awareness after compression—and outlines three approaches: dual-track position scales, document-wise coordinate resets, and Kimi K3's removal of positional encoding entirely. That last part is the real headache. Compression is straightforward; making sure the model still knows what came before what after compression is the hard part.

This one's worth opening because it turns the February 2025 bar exam disaster into law. That exam had 23 AI-generated questions slipped in without telling anyone—not the exam committee, not the California Supreme Court, not the test-takers. AB 1651's response is blunt: if AI touches the questions, the State Bar must post a disclosure 60 days before the exam. Human review doesn't erase the obligation. I'd discount the reach a bit. The law only covers materials the State Bar itself develops—contractor-delivered questions like Kaplan's 100 items are out of scope, because the Bar argued it can't know what tools contractors use internally. And there are zero penalties. No fines, no private right of action. Enforcement is basically honor system. The sharp contrast: practicing lawyers can use AI without telling clients, as long as they verify the output. But for the licensing exam that lets you become a lawyer, human review can't wash off the AI label. The entry gate is stricter than the profession itself.