all posts

The numbers are what make this worth clicking: 90% of inbound leads handled by AI, 93% of support cases resolved, combined annual infra cost in the single-digit thousands, and a claimed 32x ROI on the SDR agent. The inbound SDR team shrank from 10 people to 1.25. I'd discount this a bit. It's Vercel's own data, and inbound is a highly structured use case—leads are already coming in, no cold outreach required. The COO herself says the bottleneck isn't the model, it's codifying the sales workflow. That means this won't translate cleanly to complex outbound or enterprise sales. Still, even if you halve the claimed ROI, replacing most of an inbound team for a few thousand dollars a year is a number that sticks with anyone running a SaaS budget. The post doesn't disclose model choice or latency, so I'll wait for the October 1st Office Hours for more detail.

This piece flips a default assumption: you don't need frontier models to raise €3B and lock in Airbus, ASML, and the French Ministry of Defense. Mistral Large 3 scores below the median on Artificial Analysis—roughly on par with OpenAI models from 18 months ago—but its customers don't care. They're buying data locality, legal accountability, and supply-chain independence. After the US restricted two Anthropic models from overseas access in June 2026, that pitch got a lot sharper. I'd discount the revenue numbers a bit. They're all self-reported, unaudited, and the French entity's financials aren't publicly available. The CEO said 2026 chip and infra spend will roughly match expected revenue, and the company carries $830M in debt. Mistral proved you can raise capital and win government deals without frontier models; it hasn't proved you can turn a profit. Don't over-rotate on the market size either. Gartner's €126B European sovereign cloud estimate mostly flows to cloud providers, not model vendors. Cohere's ~$240M annualized revenue is already top-tier for this playbook. Aleph Alpha got absorbed by Cohere, and AI21 Labs laid off ~61% of staff. The sovereign AI market is real but narrow—Mistral's valuation is unlikely to repeat elsewhere.

These three sit together because they all point the same way—making models more useful with less human babysitting and fewer data leaks—but each is missing key pieces. Runway Solaris is the flashiest, and I'd discount it the most. It generates interactive UIs frame-by-frame with no frontend code: clicks and drags drive the next frame directly. Runway's own study says users prefer it 61% to 24% over Claude Opus 5-coded interfaces on instruction-following, and 71% to 21% on behavioral naturalness. Cost-wise they claim it's orders of magnitude cheaper than standard video diffusion models. Both reference points are off: the preference study pits subjective "naturalness" against code, and the cost comparison is against already-expensive video diffusion, not normal web rendering. More importantly, there's no public testing, no pricing, no API—just curated demo videos and a waitlist. Reddit devs did the math: traditional frontend code, once deployed, costs nothing per click; frame-by-frame generation burns GPU on every interaction, so costs scale linearly with usage. Runway's own post lists four unsolved limits: text rendering is unstable, the model hallucinates plausible-but-wrong UI states, visual style drifts in long sessions, and pure pixel output can't work with screen readers. Those four alone block general software production. Google WikiSkill tackles a real pain point: model weights are frozen after deployment, so every hard-won debugging lesson vanishes when you close the chat window. Their system distills agent failure logs into structured skill manuals, with a gating mechanism that only merges updates if validation scores strictly beat the historical best. On Gemini-3.5-Flash, accuracy went from 49.5% to 68.1%. Two counterintuitive findings matter more: dumping raw debugging logs directly into the executing agent's context dropped scores from 63.7% to 60.9%—noise hurts attention allocation. And skills distilled from a small model (Qwen-3.5-4B) tanked a larger model's accuracy from 50.5% to 18.1%—the small model's workarounds actively constrained stronger reasoning. The paper is CC BY 4.0 but has no official code repo, only a third-party reproduction. GitHub GHES 3.22 is the most concrete and the least complete. It lets enterprises self-host Copilot CLI inside air-gapped networks, with admins managing model endpoints centrally. But it's explicitly labeled a technical preview, many features are disabled, and future versions may change things. The real catch: data only truly stays on-prem if the model endpoint is also inside the air gap—enterprises need to run their own inference infrastructure, which isn't a flip-a-switch thing.

The reason to click: the theater is hard to ignore. Jensen Huang took a call from President Trump while onstage at the All-In Summit and told the audience 'we're not going to let an AI slowdown happen.' He publicly broke with Elon Musk, Sam Altman, and Anthropic CEO Dario Amodei—all three had previously called for slowing AI development. Nvidia makes more money selling AI chips than anyone, so Huang opposing a slowdown isn't exactly a plot twist. The interesting part is that he chose to make this statement in front of Trump, at a public event. The post doesn't disclose what Trump said, how long the call lasted, or whether 'we' means Nvidia or the whole industry. I'd discount this a bit: it reads more like a staged political signal than a policy discussion.

The reason this is worth a click: it solves a real friction when AI agents generate charts. Instead of spitting out a pile of HTML, CSS, and JS that's painful to audit, dbt Charts gives the agent a single YAML file—variables, SQL queries, 16 chart types, all in one place. The CLI renders it to SVG, HTML, or PNG, and because it lives in the same repo as dbt models, ref() catches renamed models or missing columns at PR time. The team designed this explicitly for chat agents. Strict YAML and SQL validation creates a tight feedback loop, flagging problems before anyone sees the board. Over 1,100 config options suggests they're aiming for real dashboard needs, not a toy. No release timeline yet—just a GitHub repo and docs. I'd treat this as beta. The language design is clean, but the agent experience and ecosystem aren't proven. If you're already on dbt, this direction is worth tracking. If you're just curious about "AI writes charts," wait for someone to post real-world war stories.

This one matters because it draws a legal line for agent behavior. Amazon sued Perplexity over Comet, a browser whose AI assistant navigates Amazon on a user's behalf—placing orders, comparing prices. A lower court issued an injunction, ruling Perplexity likely violated the CFAA by accessing Amazon's servers without authorization. The Ninth Circuit reversed: the "access" was performed by the user, not Perplexity. The assistant is just a tool. For anyone building agent products, this is a favorable precedent for now: if a user triggers the action and the AI executes it, liability may sit with the user, not the platform. But don't read this as "agents can freely scrape any site." The ruling is narrow—it only addresses the CFAA's definition of "access." Amazon can still pursue contract, trademark, or unfair competition claims. The case was remanded, so it's not over.

I clicked on this because Litt's stance is refreshingly direct: stop arguing about whether AI can win an IMO gold medal—that ship has sailed. He assumes superhuman math AI is imminent and asks the harder question: if high-quality math results cost a few dollars to generate, what's the point of journals and peer review? His answer isn't "reform the journals." It's to protect the things that actually build human understanding: learning seminars, serendipitous conversations, students dropping by to talk math. This isn't another "education in the AI era" piece with curriculum proposals or tool recommendations—it's about human interaction and community. I'd discount it slightly: the post doesn't lay out concrete institutional designs, so it reads more like a position statement. But the framework—protect understanding, not output—is sharper than most AI-and-academia discussions. If you care about how research communities hold together when automation arrives, this is worth ten minutes.

This one's worth opening because it's a concrete engineering story: Anthropic's team started using Claude Code heavily, PR volume spiked, and CI couldn't keep up. Instead of throttling agent commits, they rewrote the test impact analysis service. The trick is caching file dependency graphs with Merkle trees so only truly affected tests run—latency dropped from 11 seconds to under 1 second, handling 1,000 analyses per day. This is their internal monorepo setup, so direct portability varies, but the caching strategy and API design are solid references. I'd read this as a case study for rethinking CI in the agent era, not a drop-in solution.

The headline grabs you: training 1000-layer nets without backprop. Sakana AI adds dual neurons per layer—think of them as tiny PI controllers—so each layer's local dynamics can propagate supervision signals without the forward-then-backward lockstep. That fixes the signal decay that kills standard predictive coding in deep nets. I'd discount this a bit. The tests are on Fashion-MNIST and CIFAR-10 with residual MLPs, not Transformers or modern CNNs. They say performance "nearly matches" backprop but don't give the exact gap. Code and paper are public, so if you're into neuromorphic hardware or biologically plausible learning, it's worth a look. For now, this is a proof of concept, not a backprop replacement.

This is worth opening because Apple finally shipped Siri AI to users. It's not just a voice assistant upgrade—it's system-level integration: personal context across messages, mail, and photos; onscreen awareness; and camera input that can, for example, split a bill and pay via Apple Cash. Conversation history syncs across iPhone, Mac, iPad, and Vision Pro via iCloud. I'd discount the depth, though. This is an Apple press release, not a technical post. No mention of the underlying model, on-device vs. cloud, latency numbers, or privacy architecture. English public beta is live now; French, Japanese, Korean, Portuguese, and Spanish arrive in October. No word on Chinese support. If you're tracking Apple Intelligence rollout timing, this confirms the date and feature scope. Technical details will need teardown reports.

The $4.65B seed round at a $5B valuation is the headline, but the early demo is what makes this worth a click: their system beat humans and existing agents on GPU kernel optimization in under two days, without CUDA experts on the team. I'd discount this a bit — we only have Socher's verbal claims from the podcast and a 20-minute keynote summary. No paper, no benchmark, no reproducible eval yet. The vision is recursive self-improvement compressing AI research from 'thousands of people over years' into weeks, but the post doesn't spell out the system architecture or how it generalizes beyond this narrow task. His critiques of Anthropic-style constitutions and his framing of open-source as geopolitical soft power are actually more concrete than the demo itself. Don't read this as 'superintelligence is here.' A top NLP veteran just raised a massive round to bet on automated scientific discovery, and the first signal is strong but narrow. A lot of pieces are still missing.

The numbers here are hard to ignore: $11.5B in quarterly revenue, 14x year-over-year, a $65B annualized run rate, and a possible $2T+ IPO valuation. But I'd discount the 'second straight profitable quarter' claim—it's an adjusted metric that leaves out stock-based compensation, which is real dilution for future shareholders. The 80%+ gross margin is also pre-revenue-share with Amazon and pre-training cost, so it's not what you'd normally call gross margin. SemiAnalysis expects investors to target $120B annualized by year-end and nearly triple that by end of 2027, which is an aggressive bet even by AI standards. Dario calling for a slowdown, Altman and Musk backing it, and Altman saying OpenAI won't IPO this year—read together, this looks like positioning ahead of a mega-listing rather than a sudden policy awakening.

Maksim Ilin pulled together LinkedIn, WEF, Stanford, and Bain data for a September 2026 snapshot of the AI labor market. The numbers worth noting: a 62% wage premium for AI skills in the US, median Research Scientist pay of $746K–$1.15M at frontier labs, and Forward Deployed Engineer postings up over 1,000% YoY. I'd discount this a bit — it's a personal blog aggregating public data, not original research — but the sources are solid enough to use as a temperature check. A few signals feel real: Prompt Engineer has dissolved as a title, with the skill absorbed into other roles. Over half of AI postings sit outside IT departments, which suggests adoption is moving into business lines. And entry-level hiring for 22–25-year-olds trails the rest by 19% — getting a foot in the door is genuinely harder now.

Tom Gally took Simon Willison's classic pelican-riding-a-bicycle test and scaled it to 30 SVG drawing prompts. The 2026 run covers six models: GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash, DeepSeek V4 Pro, Qwen3.8 Max, and Fugu Ultra v2. No subjective ratings—just raw SVG outputs side by side so you can compare directly. The time and cost numbers are the useful bit. DeepSeek V4 Pro finished one image in 1 min 36 s at $0.10. Qwen3.8 Max took over 12 minutes—nearly 8x slower. Fugu Ultra v2 was the priciest at $1.02 per image. But remember: this is SVG code generation, not actual image generation. Don't read it as a vision benchmark. I'd discount this a bit for now. Only the first 10 prompts are done for 2026, so the sample is small. Also, the site was built by Claude Fable 5.1 and the text was AI-written—Tom just paid for it. Honest disclosure, but it means this is a lightweight experiment, not a formal eval.

The headline number is big, but the signal is in the usage: OpenAI's Temporal consumption grew 60x in under a year, and Snap runs 414 million Stories a day through it. Temporal's pitch is Durable Execution—write normal code, and the platform preserves state and auto-recovers from failures so an agent can run for days or months without hand-rolled retry logic. Annualized revenue is up over 200% YoY, net dollar retention above 200%, and they processed 1.9 trillion billable actions in August. Those aren't vanity metrics; they say reliability for long-running agents is becoming table stakes, not a differentiator. I'd discount the post a bit because it doesn't spell out how the $550M will be deployed beyond global growth. But the valuation itself tells you the market now treats agent infrastructure as a core layer, not a side bet.

This one's worth opening because Apple is finally loosening Siri's plumbing. Code sleuth 'pdfu' found references to third-party model backends in iOS 27 and macOS Golden Gate private frameworks — the direction is clear: system-level hooks for ChatGPT and Claude. Two things the post doesn't spell out. First, scope: is this a full Siri backend swap, or only for specific intents like email drafting or knowledge queries? The code doesn't distinguish. Second, timeline — no beta signs yet, and code existing doesn't mean it ships with iOS 27. I'd discount this a bit. Apple already plugged ChatGPT into iOS 18, but that was a "Siri can't handle this, want me to ask ChatGPT?" prompt — not the silent swap these hooks suggest. If this lands as a quiet replacement, it means Apple's admitting its own models can't keep up on general QA and is willing to hand over the entry point. That signal matters more than the code itself.