Key Highlights

  • Stanford’s updated “Canaries in the Coal Mine” finds entry-level employment in AI-exposed jobs down ~11% for workers aged 22–25 since 2022 — while older workers in the same occupations are largely untouched. Lead researcher Erik Brynjolfsson: “The entry-level effects we’re measuring are real, persistent and widening.” The mechanism is reduced hiring, not firing.
  • Dylan Patel’s number of the day: by the end of 2028, two labs could control ~100 gigawatts between them — most of the usable compute on Earth. His own framing of the risk is blunt: if those models are misaligned, “most of the world is misaligned basically because most of the world’s minds are there.”
  • The Ox Alpha mystery is solved: it’s Z.ai (Zhipu), and the weights are coming. The stealth model became the biggest launch in OpenRouter’s history, more than doubling DeepSeek’s usage.
  • OpenAI published the first benchmarks for Jalapeño, its own inference chip — claiming up to 3.6x faster responses and 1.9x better efficiency per watt than Nvidia’s flagship. It won’t be sold to anyone.
  • The counter-current to all that centralization shipped today too: Apple put a 512GB, 1.2TB/s M5 Ultra in the Mac Studio, explicitly optimized for people daisy-chaining desktops to run open-weight models locally.

Analysis & Opinion

AI is hitting entry-level jobs hardest, Stanford study finds — Ars Technica (via Hacker News)

The August 2026 revision of “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence” is the most careful evidence yet that the labor-market effect of AI is real, narrow, and aimed squarely at the bottom of the ladder. Working from a large subsample of anonymized high-frequency ADP payroll data, the Stanford economists found essentially no economy-wide difference in employment between the most and least AI-exposed occupations — the aggregate story that reassures everyone is basically true. Isolate workers aged 22 to 25, though, and employment in the top 40% of AI-impacted jobs has fallen roughly 11% since 2022. The mechanism matters enormously for how you’d respond: the damage shows up as lower hiring rates for entry-level roles rather than increased firings or quits, and it lands on employment levels rather than pay. The paper’s sharpest analytical move is borrowing Anthropic’s Economic Index distinction between automative use (AI replacing a human’s work) and augmentative use (AI making a human better at work they still do) — and finding that the split predicts outcomes: “The findings are consistent with automation-oriented uses of AI substituting for labor while complementary uses are associated with flat or rising employment.” Using O*NET’s required-education levels as a proxy, they also find that occupations built on codified knowledge — the formal, documented, teachable-from-a-textbook kind — show slower entry-level growth, while tacit-knowledge occupations grow faster. Read that next to yesterday’s Lars Faye piece on the “expert novice” and the trap gets much worse than either article alone: the codified-knowledge jobs disappearing are precisely the apprenticeship rungs where tacit expertise used to get built.

Google at the Global Forum on Intellectual Property — Google

Kent Walker, Google and Alphabet’s president of global affairs, used an IP Week keynote in Singapore — “The breakthrough in how we make breakthroughs” — to argue the AI-and-copyright question forward rather than defensively. His efficiency claim is the eye-catching one: recent models are on the order of 300 times more capable than those of two years ago, and today’s systems don’t merely predict but operate independently across environments and self-correct when they hit obstacles. The policy frame he’s selling is a both-and: protect creators’ rights while keeping training viable, with AI-for-science as the proof that the second half is worth defending — he points to Oxford researchers using AlphaFold to design better malaria therapies. It’s worth reading skeptically, since Google is among the most exposed parties in every training-data suit currently filed, and “we must protect creators” from the company being sued by creators is a position with obvious instrumental value. Still, it lands in a genuinely unsettled moment: the Alsup ruling covered in this digest earlier in the week suggested courts may care far more about how the corpus was acquired than whether training is transformative, which is exactly the terrain Walker is trying to shape.

The full stack behind abundant intelligence — OpenAI

CFO Sarah Friar’s argument is that gains across chips, compute, models, and products compound rather than merely add — each layer’s improvement lowers the cost of the next. Published the same morning as the Jalapeño benchmarks, it reads as the strategic thesis for why OpenAI is building its own silicon: vertical integration is the compounding mechanism, not a side quest. Dylan Patel’s podcast below is the useful adversarial read — he argues the labs will increasingly point that compounding inward, at training, rather than passing it to customers as cheaper tokens.

We’re partnering with the State of Delaware to provide free AI and career training — Google

Google is handing Delaware residents no-cost licenses for its AI courses (including the Google AI Professional Certificate) and its Career Certificates in cybersecurity, data analytics, digital marketing, IT support, project management, and UX design, administered through the state’s Department of Labor and its public libraries. The explicit design choice is that no college degree or previous experience is required, with three-to-six-month part-time completion and an employer consortium of 150+ US companies including Deloitte, Siemens, and Wells Fargo; Google cites 70%+ of Career Certificate graduates reporting a positive career outcome within six months. Set against the Stanford finding above, the program is aimed at exactly the right population and is also, unavoidably, small relative to the problem — and worth noting that several of those certificate tracks (IT support, data analytics) sit in the codified-knowledge category the Stanford paper flags as most exposed. The interesting tension is that credentialing is one of the few buffers the researchers did find: occupations with more college graduates showed “muted differences” between exposed and unexposed work.

A Manifesto for Responsible Agentic Coding — techwerkers.nl (via Lobsters)

A deliberate search for the middle ground “between mindlessly giving in to the hype created by big tech and entirely boycotting generative AI,” and its central commitment is accountability rather than abstinence: users deserve “confidence that the software they use has been reviewed and understood by humans who can take informed accountability for it.” The load-bearing principle is one that cuts against the whole generate-more-code posture — “every line of code is a liability that will have to be maintained over time” — plus a call for transparency about where LLMs were involved. It’s the same structural conclusion Theo reached from the other direction yesterday: the constraint worth optimizing isn’t generation speed, it’s verifiability.

How Much of HN is AI? — coredump.cx (via Hacker News)

An attempt to actually measure the thing everyone complains about, by sampling the daily top-5 across a full month rather than trusting vibes. The author’s framing is honest about his own conflicted position — Hacker News is both his largest traffic source and a reliable supplier of insults — and he distinguishes this moment from the 2018 crypto/NFT wave as something more pervasive: stories about AI, written by AI, and commented on by AI.

OpenAI loses a top data center exec as stream of high-profile departures continues — TechCrunch

Chris Malone, OpenAI’s head of data centers, left last week per the Wall Street Journal, after a relatively short tenure that began in March of last year — following nearly five years at Meta and over a decade at Google. The timing is what makes it notable rather than routine: he owned execution of the buildout underpinning the Stargate Project, at a moment when data center strategy is arguably the most consequential role at any frontier lab.

New Products & Tools

Jalapeño’s first results show industry-leading speed and efficiency in AI inference — OpenAI

OpenAI took its first custom silicon to Hot Chips with benchmarks attached, and the claims are aggressive: measured on SemiAnalysis’s InferenceX benchmark, the 700-watt part delivered both more tokens per user and more throughput per kilowatt than current state-of-the-art inference processors — per The Rundown, up to 3.6x faster responses at 1.9x the efficiency per watt versus Nvidia’s flagship. Hardware chief Richard Ho’s framing covers both axes that usually trade off: “It’s very efficient to serve a lot of customers, but it can also be very low latency.” Two details matter more than the benchmark. First, OpenAI won’t sell it — Ho says the company “has so much need for it” internally — and it still buys Nvidia for training, so this is a margin play on inference, not a challenge to Nvidia’s core franchise. Second, OpenAI used its own Astra model and Codex to design the chip, going from concept to manufacturing-ready in nine months; two further generations are already in development, landing in data centers late 2026 with production ramping through 2027. Also see TechCrunch’s benchmark writeup — the chip was first announced last October and built with Broadcom.

China’s Z.AI made Ox Alpha, the stealth model that rivals DeepSeek — Bloomberg (via Hacker News)

The guessing game this digest covered on Monday is over: Z.AI (Zhipu) confirmed Ox Alpha is a new iteration of its GLM series, and said it will release the weights. The adoption numbers justify the attention — it went straight to the top of OpenRouter’s leaderboard, more than doubling DeepSeek’s usage, in what the marketplace calls the biggest launch in its history; Stripe CEO Patrick Collison (whose company is acquiring OpenRouter) called it “very impressive.” It’s a reasoning model built for coding and agentic work, handling text, image, and video input, free for its first week with pricing still unannounced. The stealth-launch playbook — used before by Alibaba and Xiaomi — lets a lab collect real usage data before attaching its name to the result, and the code name references the recent Chinese film Ox Comes. Note how this reframes Dylan Patel’s compute argument below: Z.ai is reportedly running a fraction of the frontier labs’ compute and still landing near the frontier, which is the strongest available case that the centralization thesis isn’t destiny.

With new Mac Studio and Mac mini, Apple leans hard into local AI inference — Ars Technica (via Lobsters)

A spec bump whose positioning is the actual news: Apple introduced the M6 (its first 2nm Mac chip, 12-core CPU with two new “super cores,” but capped at 32GB and 160GB/s) and the M5 Ultra, effectively two M5 Maxes on one package — 36 CPU cores, 80 GPU cores, up to 512GB unified memory and 1.2TB/s of bandwidth. The reason those numbers exist is a use case Apple didn’t design the original machines for: since macOS 26.2 enabled low-latency Thunderbolt 5 host-to-host communication for distributed MLX inference, developers have been daisy-chaining minis and Studios to serve open-weight Qwen and DeepSeek models far larger than any single consumer device can hold. Mac mini starts at $899 (M6) and $1,699 (M5 Pro); Mac Studio at $2,499 (M5 Max) and $5,499 (M5 Ultra), shipping September 22 with macOS 27 Golden Gate.

Claude Cowork finally remembers what you told the app in chat — TechCrunch

Anthropic merged the memory systems behind Claude chat and Claude Cowork, so context established while thinking through a problem carries into the session where you act on it — and it’s exposing stored memories for users to read, edit, or delete. Worth reading against Theo’s audit yesterday, which found a 3:1 write-to-read ratio on Claude Code memories: user-visible, user-editable memory is the direct answer to his complaint that the failure mode is silent accumulation of stale facts.

Introducing the Admin plugin for ChatGPT Work and Codex — OpenAI

An admin surface for analyzing workspace usage, managing members and permissions, adjusting limits, and acting on admin requests — aimed squarely at the enterprise-governance friction that OpenAI’s own product leadership has named as a brake on Codex adoption.

Accel-backed Keenable is indexing the web for AI agents — TechCrunch

Andrey Styskin (formerly head of search, AI, and cloud at Yandex) and Matthias Petri came out of stealth with $26M seed and an index of 100 billion+ documents, on the thesis that search infrastructure built for humans who won’t read whole pages is the wrong shape for agents. The API is reportedly already in production at several AI labs and inference providers, used during both training and runtime.

Robotics startup Generalist reaches $3B valuation — TechCrunch

Nearly $200M more led by 8VC extends June’s $400M Series B and lifts the valuation from $2B to $3B. Founded in 2024 by ex-DeepMind researchers Pete Florence and Andy Zeng with former Boston Dynamics engineer Andrew Barry, it claims its new Gen 1.5 model lets robots pick up tasks from video demonstrations as short as 3 to 12 seconds.

Stability AI raises $76 million in fresh funding — TechCrunch

A $76M Series B takes Stability’s total to $232M, and the cap table is the story: Universal Music Group, Sony Music Group, Warner Music Group, and Electronic Arts alongside AMD Ventures and Pacific Alliance Ventures. The company that once defined adversarial-to-rightsholders generative AI is now substantially funded by rightsholders.

Runable hits $21M to bet AI agents can go from building businesses to growing them — TechCrunch

The Indian startup raised a $21M Series A co-led by Susquehanna Venture Capital and Nexus Venture Partners at a $65M post-money, wagering that the scarce thing is no longer building an app but finding customers for it. CEO Umesh Kumar’s pitch is a useful corrective to tool-centric framing: “In the end, a business doesn’t require Codex or Claude Code or anything. They require real outcomes.”

India’s Ringg gets backing from Peak XV as it pushes voice AI past the phone call — TechCrunch

Ringg took $10M from Peak XV as a Series A extension (total round $15.5M) while running 20 million call attempts monthly, against Truecaller data showing 76%+ of Indian consumers prefer phone calls when contacting businesses. It pivoted from text-to-speech platform DesiVocal after concluding model training was too expensive and the use cases weren’t sticky.

NVIDIA announces Jetson Orin Nano 2 — NVIDIA

The entry-level edge robotics board doubles inference performance in the same form factor at 40% less power for equivalent work, with 78 TOPS, 8GB of memory, and an 8-core Arm CPU — aimed at robots, delivery drones, and inspection systems.

CUDA Python 1.0: Stable APIs, One Foundation, Full Platform Access — NVIDIA Developer

Shipping alongside CUDA 13.3, this consolidates years of fragmented GPU access from Python — where PyTorch, CuPy, and RAPIDS each reached CUDA their own way and interoperating on shared data was painful — into one NVIDIA-maintained stack: cuda.core, cuda.compute, cuda.bindings, cuda-pathfinder, and nvmath-python 1.0.

Leading publishers bring blockbuster PC games and technology to NVIDIA RTX Spark — NVIDIA

EA, Embark, and Ubisoft committed support for RTX Spark’s fall launch, a Windows platform bundling personal AI agents and content creation with gaming; announcements included DLSS 4.5 Ray Reconstruction and native Javelin Anticheat integration across EA’s multiplayer titles.

IMDEX uses Cursor to build integrated subsurface data platform in months, not years — Cursor

The mining-technology firm reports finishing a data-platform consolidation in eight months — work it estimates would otherwise have needed 20 additional engineers and multiple years — and an Angular-to-React migration in six months against a three-to-five-year projection. As with all vendor case studies, the counterfactuals are self-reported.

Gamma acquires Accel-backed design startup Lica — TechCrunch

Gamma is turning Lica into its design research lab, with co-founders Priyaa Kalyanaraman and Purvanshi Mehta leading it, as the presentation company pushes beyond slides into other visual communication formats.

How loveholidays is making everyone a builder with Codex — OpenAI

A customer story on using Codex to move software development beyond the engineering org, letting non-engineers turn ideas into shipped product features.

Show HN: CarWatch — a Raspberry Pi with Qwen as a local car AI — GitHub (via Hacker News)

A Raspberry Pi 5 running Qwen3.6-35B-A3B entirely offline at ~3.5 tokens/second and ~65°C sustained, doing local RAG over the vehicle’s 745-page owner’s manual with page citations. The design constraints are the point: works with no signal in garages and tunnels, vendor-neutral across car brands, and no data leaves the vehicle.

You can officially buy the Pixel 11 phones and Pixel Watch 5 — Google

The Pixel 11 line (standard, Pro, Pro XL, Pro Fold) is on sale from $899 to $1,899 with Gemini Intelligence features throughout, alongside Pixel Watch 5 at $399/$429 with new Health Guardian emergency detection. Google also documented intelligent dictation in Gemini for macOS.

GitHub Trending — GitHub

Agent skills and plugin marketplaces have taken over the board. freestylefly/awesome-gpt-image-2 is running away with the day (+4,044, 20.5k stars) on a prompt-engineering template library; DietrichGebert/ponytail follows (+1,598, 111.9k) with the best pitch on the page — “makes your AI agent think like the laziest senior dev in the room.” Then MadsLorentzen/ai-job-search (+1,299, 36.0k), Alishahryar1/free-claude-code (+1,047, 50.1k), basecamp/omarchy (+1,021, 31.7k), and tt-a1i/archify (+1,002, 16.8k), a skill for generating verifiable architecture diagrams. Also climbing: rohitg00/ai-engineering-from-scratch (+837), AgriciDaniel/claude-obsidian (+812), anthropics/claude-plugins-community (+537 on a 2.0k base), tinyhumansai/openhuman (+522), VoltAgent/awesome-agent-skills (+488, now cataloguing 1,000+ skills), marin-community/marin (+443), and anthropics/claude-plugins-official (+307, 34.2k).

Research

Inside LinkedIn’s cognitive memory agent for agentic personalization — Stack Overflow Blog

Principal AI researcher Praveen Bodigutla walks through the four-layer memory architecture behind LinkedIn’s hiring assistant — conversational memory for recent turns, semantic memory aggregating a recruiter across sessions and surfaces, and further layers above — built on the observation that recruiters’ expressed hiring preferences stay consistent across similar roles. The design goal is persistent personalization rather than per-session context, which is the same problem Anthropic is attacking from the product side with today’s Claude memory merge.

Restore LLM inference capacity in seconds with shadow engine recovery in NVIDIA Dynamo — NVIDIA Developer

Cold-restarting a failed LLM engine means reloading weights into HBM, recompiling kernels, and recapturing CUDA graphs — minutes during which surviving workers absorb the displaced traffic. Dynamo’s preview feature keeps a fully initialized shadow engine idle on the same GPUs, sharing weights via the GPU Memory Service so there’s no second HBM copy, and cuts failover to seconds.

Super-intelligence or Superstition? Exploring Psychological Factors Influencing Belief in AI Predictions about Personal Behavior — arXiv (via Lobsters)

Across 238 participants, Eunhae Lee, Pat Pataranutaporn, Judith Amores, and Pattie Maes found that belief in AI predictions correlates positively with belief in astrology and personality-based prediction, and that paranormal beliefs plus favorable AI attitudes both raised perceived validity. Conscientiousness predicted less acceptance across all prediction sources, while cognitive style had no significant effect — a pattern the authors call “rational superstition,” where “belief is driven more by mental heuristics and intuition than critical evaluation.”

AI At Home Part 2: Multi GPU Drifting — jdagostino.github.io (via Lobsters)

A follow-up on squeezing throughput out of a home server built from budget GPUs, focused on applying existing techniques rather than novel ones, with a from-first-principles walkthrough of transformer inference and tokenization for readers arriving cold. A useful hobbyist counterpart to the Apple local-inference story above.

NVIDIA’s agentic-infrastructure papers — performance per watt, Spectrum-X Ethernet, and DSX MaxLPS — NVIDIA Developer

Three technical companions to the Vera Rubin launch covered earlier this week. The recurring datum is from OpenRouter’s analysis of 100 trillion real tokens — average prompt tokens per request up roughly 4x, with a single agentic request consuming 15x the tokens of a standard chat turn — which is the justification for benchmarking on replayed production sessions (SemiAnalysis AgentX) instead of synthetic workloads. The MaxLPS piece supplies the uncomfortable efficiency number: in representative deployments only about 60% of delivered site power reaches compute, with the rest going to cooling, networking, storage, and facility overhead, made worse by static rack provisioning for worst-case draw.

Interviews & Conversations

Dylan Patel – Two labs will soon control most of the world’s workforce — Dwarkesh Patel (76:53)

Transcript-based summary. The SemiAnalysis founder lays out an economic case for compute centralization that is more specific, and more unsettling, than the usual version. The hinge is unit economics: serving GPT-4 on Hopper generated negative gross margin for OpenAI, but a megawatt now costs roughly $10–15M to stand up while Anthropic is extracting as much as $50M of revenue per megawatt — so every dollar of inference capacity throws off several dollars that get plowed straight back into training. Anthropic turned a profit in Q2 and Patel thinks OpenAI may have in Q3, meaning the buildout is no longer venture-funded. From there the ratios compound: the two labs went from ~2GW each in January to above 5GW by year-end, took ~30% of all incremental compute this year, have 40–50% of next year’s already contracted, and could hold ~100GW combined by end-2028 — against maybe 200GW worldwide, and at a higher performance-per-watt than everything built before it. His most contrarian claim is that the labs will allocate a decreasing share of compute to inference, against the consensus that inference dominates; he argues it’s already happening, since Anthropic’s monthly compute additions keep rising while its revenue additions have plateaued. The logic is brutal and hard to refute: if you can turn a megawatt into $100M of revenue or point it at building AGI, the board picks AGI.

Two things bend the curve. The first is finance. SemiAnalysis models $11 trillion of capex from 2024–2029, roughly $6T from cash flows and $5T raised as debt — enough, Patel argues, to push rates up across the entire economy. If Meta moves from ~5–6% to 8% because the return on compute justifies it, everyone else pays that 250bp too: banks get squeezed as liabilities reprice faster than assets, and every long-duration cash-flow equity — his examples are Johnson & Johnson and railways — reprices downward even as the S&P holds. Dwarkesh extends it to sovereign debt, noting that ~80% of US federal revenue is payroll and income tax (the automatable kind) while corporate income tax is under 10%, so a few points of rate increase could take debt service from 20% of revenue past 60%; he expects a second Volcker shock, and names Pakistan and Nigeria as the countries that default. The second brake is politics, and Patel is emphatic that it’s already binding: New York banning data centers, Texas moratoriums, Ohio floating property-tax obligations, OpenAI pausing training for two weeks, Astra not widely deployed even internally, Anthropic withholding Mythos from foreign employees. “Everyone who’s elected is going to hate AI. Even the people who are elected already hate AI.”

The part worth sitting with is the labor framing that gives the episode its title. Frontier compute is growing 4–5x a year while the compute needed for a given capability falls ~3x, so the effective population at a fixed capability level grows roughly 10x annually — no recursive self-improvement required. Extrapolate and a single lab goes from ~10 million AI workers to 100 million to a billion, and before the decade is out one company commands more labor-equivalents than there are people on Earth. Dwarkesh draws the conclusion plainly: “if these AIs are misaligned then most of the world is misaligned basically, because most of the world’s minds are there.” Patel’s answer to whether anything avoids this is not reassuring — “it seems to me that every force is screeching towards centralization, and that’s scary as hell” — and his stated hope, that the value keeps escaping to customers like Jane Street capturing $300–500M per megawatt, he labels himself: “This is my cope.” Note the two counterweights elsewhere in today’s digest: Z.ai reached near-frontier quality on a rounding error of that compute, and Apple shipped hardware for running capable models in your own house.

References

  1. “AI is hitting entry-level jobs hardest, Stanford study finds,” Ars Technica via Hacker News, 2026-08-25 [blog]
  2. “Google at the Global Forum on Intellectual Property,” Google, 2026-08-26 [blog]
  3. “The full stack behind abundant intelligence,” OpenAI, 2026-08-25 [blog]
  4. “We’re partnering with the State of Delaware to provide free AI and career training,” Google, 2026-08-25 [blog]
  5. “A Manifesto for Responsible Agentic Coding,” techwerkers.nl via Lobsters, 2026-08-25 [blog]
  6. “How Much of HN is AI?,” coredump.cx via Hacker News, 2026-08-25 [blog]
  7. “OpenAI loses a top data center exec as stream of high-profile departures continues,” TechCrunch, 2026-08-25 [blog]
  8. “Jalapeño’s first results show industry-leading speed and efficiency in AI inference,” OpenAI, 2026-08-25 [blog]
  9. “OpenAI’s Jalapeño chip is built for fast inference at scale, benchmarks show,” TechCrunch, 2026-08-25 [blog]
  10. “OpenAI’s first AI chip brings the heat,” The Rundown, 2026-08-26 [blog]
  11. “China’s Z.AI made Ox Alpha stealth model that rivals DeepSeek,” Bloomberg via Hacker News, 2026-08-26 [blog]
  12. “With new Mac Studio and Mac mini, Apple leans hard into local AI inference,” Ars Technica via Lobsters, 2026-08-25 [blog]
  13. “Claude Cowork finally remembers what you told the app in chat,” TechCrunch, 2026-08-25 [blog]
  14. “Introducing the Admin plugin for ChatGPT Work and Codex,” OpenAI, 2026-08-25 [blog]
  15. “Accel-backed Keenable is indexing the web for AI agents,” TechCrunch, 2026-08-25 [blog]
  16. “Robotics startup Generalist reaches $3B valuation, sources say,” TechCrunch, 2026-08-25 [blog]
  17. “Stability AI, maker of image generator Stable Diffusion, raises $76 million in fresh funding,” TechCrunch, 2026-08-25 [blog]
  18. “Runable hits $21M to bet AI agents can go from building businesses to growing them,” TechCrunch, 2026-08-26 [blog]
  19. “India’s Ringg gets backing from Peak XV as it pushes voice AI past the phone call,” TechCrunch, 2026-08-25 [blog]
  20. “NVIDIA Announces Jetson Orin Nano 2 Robotics Computer to Redefine Entry-Level Edge AI,” NVIDIA Newsroom, 2026-08-25 [blog]
  21. “CUDA Python 1.0: Stable APIs, One Foundation, Full Platform Access,” NVIDIA Developer Blog, 2026-08-25 [blog]
  22. “Leading Publishers Bring Blockbuster PC Games and Technology to NVIDIA RTX Spark,” NVIDIA Blog, 2026-08-25 [blog]
  23. “IMDEX uses Cursor to build integrated subsurface data and analytics platform in months, not years,” Cursor, 2026-08-25 [blog]
  24. “Gamma acquires Accel-backed design startup Lica,” TechCrunch, 2026-08-25 [blog]
  25. “How loveholidays is making everyone a builder with Codex,” OpenAI, 2026-08-26 [blog]
  26. “Show HN: I made a Raspberry with Qwen my local car AI (CarWatch),” GitHub via Hacker News, 2026-08-25 [blog]
  27. “You can officially buy the Pixel 11 phones and Pixel Watch 5,” Google, 2026-08-25 [blog]
  28. “Here’s how to use intelligent dictation in Gemini for macOS,” Google, 2026-08-25 [blog]
  29. “GitHub Trending,” GitHub, 2026-08-26 [blog]
  30. “Inside LinkedIn’s cognitive memory agent for agentic personalization,” Stack Overflow Blog, 2026-08-25 [blog]
  31. “Restore LLM Inference Capacity in Seconds with Shadow Engine Recovery in NVIDIA Dynamo,” NVIDIA Developer Blog, 2026-08-25 [blog]
  32. Eunhae Lee, Pat Pataranutaporn, Judith Amores, Pattie Maes, “Super-intelligence or Superstition? Exploring Psychological Factors Influencing Belief in AI Predictions about Personal Behavior,” arXiv via Lobsters, 2026-08-25 [blog]
  33. “AI At Home Part 2: Multi GPU Drifting,” jdagostino.github.io via Lobsters, 2026-08-25 [blog]
  34. “NVIDIA Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per Watt,” NVIDIA Developer Blog, 2026-08-24 [blog]
  35. “Giga-Scale AI and the Ethernet Evolution: How Spectrum-X Ethernet Rewrites the Rules,” NVIDIA Developer Blog, 2026-08-24 [blog]
  36. “Maximizing AI Factory Performance per Watt with NVIDIA DSX MaxLPS,” NVIDIA Developer Blog, 2026-08-24 [blog]
  37. Dwarkesh Patel, “Dylan Patel – Two labs will soon control most of the world’s workforce,” YouTube, 2026-08-25 [video]