curated · 2026-09-18

This one's worth opening because the timing is brutal. On the same day the NYT reports Trump is downplaying AI fears for economic reasons—likely resisting regulation—Katie Bo Lillis drops a concrete case: an AI-assisted intel report hallucinated that a Chinese ship was carrying nuclear weapons components. The US military scrambled to intercept it. It was a hallucination, and it 'almost started a war.' Marcus warned the Senate about exactly this in 2023. I'd separate the two threads. The Trump piece is a headline screenshot so far—no policy details in the post, so I wouldn't draw hard conclusions yet. The intel hallucination story has a named reporter and a specific event, which makes it the heavier signal. The point isn't 'AI makes mistakes.' It's that when a hallucination enters a military decision chain and triggers real action, the margin for error is zero. Marcus closing with 'we may not get as lucky next time' isn't hyperbole.

This is worth opening because Mollick isn't running benchmarks—he's throwing models into messy, real-world tasks to see what they actually do. He had GPT-6 Astra turn the 1977 text adventure Zork into a playable 3D action game, with the AI deciding what the white house and monsters look like and converting 'fight the troll' into an action sequence. In a second experiment, Fable 5.1 reconstructed Umberto Eco's private library from a dozen videos, shelf photos, and two catalogues—no floor plan available—placing roughly 5,000 identified books across 27,000 shelf slots. He also had Astra operate Blender and produce a 3D animated book trailer in 45 minutes. His core argument is 'capability overhang': these models can already do weeks of human work, but almost nobody taps that potential. He frames four human advantages to close the gap—deep knowledge, wide knowledge, taste, and agency. That framework is more useful than vague 'AI will replace you' talk. I'd discount this a bit: the post doesn't say when GPT-6 Astra or Fable 5.1 launched, whether they're publicly available, or give any specs. These experiments read more like a power user stress-testing frontier models, not something the average person can replicate today. But the direction is right—dig into what current models can do before obsessing over the next release.

I'd open this because Marcus lines up three incidents in a tight window: OpenAI employee accounts hijacked, Hugging Face breached, ChatGPT used to write malware. The WSJ op-ed he cites adds a sharper point—major labs see agentic products as their main post-IPO revenue, so they have little incentive to clamp down on misuse. I wouldn't read this as 'AI is about to destroy the internet.' The more useful framing: once models can autonomously execute tasks, call tools, and access external systems, attackers don't need deep technical skills to cause damage at scale. Marcus doesn't spell out concrete defenses, but he gets the priority right—the agentic attack surface is live and under-policed, and that's a more immediate problem than hypothetical superintelligence scenarios.

This one's worth opening because it moves AI code evaluation from benchmarks to an actual engineering feedback loop. Justin Cormack had AI build an S3-compatible object store—350k lines of Rust. Instead of reading every line, he ran 1,500 tests against real S3 as the oracle. The tests caught real S3 500 errors, which is a nice side effect. The practical lessons are what make this useful: chasing 100% coverage made the agent write trivial filler tests. AWS docs were often wrong or outdated, and AI was bad at extracting edge cases from them. His hard rule—fix flaky tests immediately or the agent learns to ignore failures—came from watching this at scale, not from theory. I'd read this as a testing-strategy guide for large AI-assisted projects, not as "AI can build storage systems." His core point: tests are discovery tools, not a scorecard. Add them where you're uncertain or suspicious, not where the coverage number needs a bump.

This one's worth opening because it reframes AI in security auditing: not as a bug-finding bot, but as a tool-builder that makes the audit possible in the first place. Miden zkVM uses a custom assembly language called MASM with zero existing developer tooling. Trail of Bits had Claude build a working LSP prototype in days—syntax highlighting, go-to-definition, hover docs—then added a decompiler, static analysis engine, and a Lean formal model. These tools caught an unvalidated input that let a malicious prover forge Falcon signatures and drain accounts. The Lean work produced 95 machine-checked correctness proofs. I'd discount this a bit: the post only names Claude for the LSP prototype. It doesn't say which models built the decompiler, static analysis, or Lean model. No numbers on time or token cost either. But the core insight holds—AI here isn't replacing auditors; it's translating their expertise into custom tooling that makes previously unauditable code auditable. That's a more useful story than the usual 'AI found X bugs' press release.

The reason to click: Qwen pushed simultaneous interpretation latency down to 2.3s and added three features that make it usable in real meetings, not just demos. The Interleave architecture weaves audio and text into one stream so already-translated chunks get cached — that's where the quality bump and latency drop from 2.8s come from. The practical upgrades: real-time speaker separation with voice cloning (so you know who said what), synchronized bilingual output on one screen, and long-context disambiguation that keeps names and terms consistent across a conversation. On Omnilingua-MSpeaker and FLEURS, it beats current mainstream systems on faithfulness, fluency, and conciseness. I'd discount this a bit: 2.3s is average latency — the post doesn't break out performance on very long sentences or noisy environments. DashScope pricing isn't mentioned either. But if you need live bilingual captions for cross-border meetings or streams, this is worth a test call.

This post is useful because it solves a real headache: image model pricing is a mess of tokens, megapixels, and per-image rates that make side-by-side comparison impossible. OpenRouter sent the same prompt to 20 models and read the actual billed cost, so the numbers are grounded. GPT Image 2 came out cheapest at $0.006 for a 1024×1024 PNG, and it even rendered text correctly. Gemini 3 Pro Image was the most expensive at $0.134. The hidden takeaway is OpenAI's quality parameter—it swung the same image from $0.006 to $0.211, a 35x jump, so you'd want to be sure the upgrade is worth it. Recraft V4.1 Vector outputs editable SVG at $0.08, which is more practical than pixels for logos or icons. I'd bookmark this as a pricing cheat sheet, but the post itself warns that prices shift fast and any figure older than a month is just a hint.