Key Highlights

  • Three separate studies now say the same thing: AI coding assistance blocks the formation of expertise. Lars Faye stitches together a JetBrains-cited study on novice programmers, a UPenn trial where AI-assisted students scored 17% worse, and Anthropic’s own 2026 research — all landing on “cognitive effort, and even getting painfully stuck, is likely important for fostering mastery.”
  • An AI assistant in private beta claims a “perpetual and irrevocable” license to everything it touches. TechCrunch found Instinct kept summarizing emails after disconnection, sent emails without permission, and could be phished into leaking sign-up codes.
  • OpenAI banned a Russia-origin network that used its models to prop up a fake Israel-based think tank and a “sovereignty” index praising Russia.
  • Theo audited Claude Code’s memory and turned it off across his entire fleet. The damning number: a 3:1 write-to-read ratio, with 26 of 45 memories never read once across 355 sessions.
  • NVIDIA’s Vera Rubin generation goes into full production alongside Groq 3 LPX — and SpaceX is putting a variant of it in orbit, with the first Starmind racks targeted for late 2027.

Analysis & Opinion

AI Coding will Prevent Expertise — Lars Faye (via Hacker News)

The sharpest piece of the day, and it names a real trap: the tools demand expertise to use responsibly, but they circumvent exactly the friction that produces expertise. Faye calls the result the “expert novice” — a developer told simultaneously that they’ll be left behind without AI and that getting value from AI requires the architectural taste that only comes from years of doing it the hard way. The evidence is unkind to the “personal tutor” hope: in the study JetBrains cites, heavy-AI participants “often skipped crucial planning stages” and finished with an “illusion of competence”, while the ones who did best had developed “negative expertise” — the ability to ignore unhelpful suggestions. UPenn’s 2025 study of 1,000 students found the AI-crutch group performed 17% worse than students with just a textbook, and thought they were excelling; the same study’s Tutor variant, which forced students to solve problems themselves, produced a 127% improvement in practice sessions. Faye’s framing is that this friction is a feature, not a bug — Fingerspitzengefühl, the fingertip feeling that tells you “this is probably going to cause problems,” is built only by failing. His conclusion is the uncomfortable one: the most productive learning with an AI coding tool happens when it isn’t used to generate much code at all.

Instinct’s powerful AI assistant is raising privacy and security concerns — TechCrunch

Instinct — a private-beta personal assistant from former Sierra researcher Noah Shinn — connects to your email, messaging apps, calendar, audio, location, and screen, and the terms of service ask for a lot in return: a “perpetual and irrevocable” license to access, store, reproduce, distribute, and modify user materials for AI training. The reported failures aren’t hypothetical. Testers found it continued summarizing emails after being disconnected and stored them in plain text, sent emails on users’ behalf without permission, and could be phished into surfacing sensitive material like sign-up codes. Users initially couldn’t delete their email records at all; a deletion tool was added only after complaints. The critics are not the usual suspects — they include Michael Mignano of Union Square Ventures, Moxxie Ventures founder Katie Jacobs Stanton, and Hello Patient’s Alex Cohen. One security expert’s summary is the whole story in a sentence: “We’re trading privacy and control for hyper-personalized AI tools…often without fully understanding the trade.” Read this one next to the OpenAI product interview below, where granting an agent access to your email and iMessages is framed as the natural next step.

OpenAI is building AI agents for everything. Will everyone use them? — TechCrunch

The adoption numbers undercut the narrative. 98% of OpenAI’s own employees use Codex, but only 17% of organizational subscribers and under 1% of individual subscribers touch the agentic coding tool — and ChatGPT Work sits at 20 million users against ChatGPT’s billion-plus. The piece is honest about why: permission settings are confusing, agents work best on measurable data-heavy tasks and flounder where success has no clear metric, and the author burned $65 of tokens in four casual days on a $20/month plan. Mario Zechner supplies the structural objection — “Everything is coding agent shaped…you cannot capture [non-digital workflows] in a simple trace” — while vertical rivals like Harvey and Clay hold the domain expertise OpenAI doesn’t.

‘The world seems to be ready’: An interview with OpenAI head of product Thibault Sottiaux — TechCrunch

Sottiaux’s product philosophy is model-first and deliberately thin: “You need to get out of the way, almost, of the model, and just let it express itself.” He points to ChatGPT Work’s 20 million users as proof that “the world seems to be ready,” and frames Luna’s pricing — an 80% cost reduction on frontier capability — as the mechanism for broadening access rather than a margin play. On the obvious objection, granting an agent access to email and iMessages, he leans on safety infrastructure and honest benchmarking as the precondition for adoption.

Responsible AI adoption needs developer workflow design — Stack Overflow Blog

Shadow AI isn’t a compliance failure, it’s a workflow signal — you don’t fix it with a policy document nobody reads, you fix it by making the sanctioned path faster than the improvised one. Stack Overflow’s own data captures the tension well: 84% of developers use or plan to use AI tools, yet more distrust AI accuracy than trust it, with the top complaint being output that looks right and then eats hours in debugging. The recommendation to leadership is to treat bypassed channels as a report on friction rather than recklessness.

Situational Awareness, star AI hedge fund that nearly imploded, now being probed by the SEC — TechCrunch

The SEC has begun subpoenaing banks that traded with and financed Leopold Aschenbrenner’s AI-focused fund, which suffered catastrophic losses in the late-July 2026 AI selloff, and has directed them to preserve records. No wrongdoing has been alleged; the fund says it will cooperate fully.

The SpaceX-Nvidia partnership heads for orbit — The Rundown

SpaceX and NVIDIA made the orbital data center official, targeting the first Starmind hardware in space by Q4 2027 on a space-adapted Vera Rubin NVL72 that Musk describes as “significantly simpler, lower cost, denser and lighter.” The economics are the open question: orbital compute currently runs about 4x ground cost, and the whole bet is Musk’s claim that the ratio inverts within a few years — a direct answer to Sam Altman having called space data centers “ridiculous.” Elsewhere in the issue: Thomson Reuters spent $40M over two years building a custom legal model by retraining Alibaba’s open-source Qwen, with the latest training run costing just $450,000; and Chinese attackers have more than doubled attack volume by leaning on DeepSeek, chosen for low cost and thin guardrails.

Trump bought SpaceX shares two weeks after blockbuster IPO — TechCrunch

The president bought up to $50,000 of SpaceX on June 23, roughly two weeks post-IPO, at a time the stock had slid from its $200+ peak to the mid-$150s — and it has since fallen back to the $135 IPO price. The White House says the portfolio is managed by third parties and tracks market indexes.

New Products & Tools

NVIDIA Groq 3 LPX enters full production alongside Vera Rubin — NVIDIA

Groq 3 LPX, an inference accelerator extending the Vera Rubin platform, is now in full production, hitting 3,400 output tokens/second on Gemma 4 31B at a 100K-token context window per Artificial Analysis — roughly 4x the nearest platform. NVIDIA’s companion post names early adopters: Nebius is the first AI cloud to take Groq 3 LPX, CoreWeave has begun production deployment of Spectrum-X Multiplane networking, and SpaceXAI is adopting Vera CPUs across its Grok infrastructure — including the orbital Starmind satellite above.

NVIDIA BlueField-4 powers “Scale-In” network infrastructure — NVIDIA Developer

BlueField-4 pairs a 64-core Grace CPU with inline acceleration engines enforcing policy at up to 800 Gb/s, offloading security and data access from host CPUs so they don’t bottleneck as compute scales. NVIDIA is pitching it alongside NVLink Fusion, which lets hyperscalers plug custom XPUs into NVIDIA’s rack-scale infrastructure rather than building the surrounding platform themselves.

Hugging Face reportedly in talks to be acquired for $13B — TechCrunch

The model-sharing platform is working with banks to evaluate bids at roughly $13 billion, up from a $4.5B post-money in 2023; the acquirer is undisclosed and no deal is final. Worth noting Hugging Face previously turned down a $500M NVIDIA investment at a $7B valuation, and CEO Clem Delangue’s stated position — “we have a long-term responsibility to them” — cuts against a sale.

Valor, Point72 back General Intuition at $6B valuation — TechCrunch

General Intuition, which trains spatial-reasoning foundation models on hundreds of millions of hours of gameplay footage plus “action labels” from parent platform Medal, is raising at a $6B pre-money — weeks after a $320M round at $2.3B. Vinod Khosla calls action labels “a key part of the ’emergence of intuition’”; it’s Valor’s first AI lab bet since SpaceX.

Advancing price-performance for developers with GPT‑5.6 in Kiro — OpenAI

GPT‑5.6 is now available in Kiro for planning, building, reviewing, and testing software, positioned on price-performance.

How developers build AI for good with Gemma 4 — Google

Winners of the Gemma 4 Good Challenge, a six-week Kaggle competition that drew 1,600+ entries for impactful AI built under resource constraints — with entrants leaning on LiteRT, Cactus, Ollama, llama.cpp, and Unsloth to run on everyday hardware.

OCR It — GitHub (via Hacker News)

A Chrome extension that pins a screen region once, then lets you hotkey through a paginated document pulling text for your LLM; OCR runs fully offline via bundled Tesseract.

Celebrate 110 years of national parks with Maps, Search, and Gemini — Google

A seasonal consumer push tying AI Mode in Search to gear recommendations and trip planning, off the back of a 165% monthly rise in “hiking for beginners” queries.

GitHub Trending — GitHub

Agent tooling still owns the list. freestylefly/awesome-gpt-image-2 leads on velocity (+1,698 today, 16.5k stars), followed by MadsLorentzen/ai-job-search (+1,266, 34.7k), openai/codex (+1,183, 117.8k), basecamp/omarchy (+1,080, 30.7k), and DietrichGebert/ponytail (+944, 110.3k). Also climbing: multica-ai/andrej-karpathy-skills (+828), AgriciDaniel/claude-obsidian (+810), rohitg00/ai-engineering-from-scratch (+572), tinyhumansai/openhuman (+541), apache/maka (+538), and anthropics/claude-plugins-community (+350 on a 1.5k base).

Amjad Masad joins the Disrupt Stage at TechCrunch Disrupt 2026 — TechCrunch

Replit’s CEO will speak on programming’s future at Disrupt (Oct 13–15, Moscone West); the company’s valuation went from $3B to $9B in six months this year.

Research

Disrupting a new covert influence campaign from Russia — OpenAI

OpenAI banned a cluster of Russia-origin accounts using its models to manufacture the appearance of independent geopolitical analysis. The operation’s core assets were a fabricated Israel-based think tank and a “sovereignty” index engineered to rank Russia favorably while criticizing the West — a laundering pattern that matters more than raw content volume, because the credibility comes from the fake institution rather than the text. It’s a useful counterweight to the framing that AI influence operations are mainly a scale problem; here the leverage is in cheap institutional legitimacy, not cheap words. Reports like this remain one of the few windows outsiders get into what model providers actually catch, and how.

Up to 30x more work per watt: Vera Rubin NVL72 sets a new efficiency standard for AI agents — NVIDIA

Measured on the SemiAnalysis AgentX benchmark, Vera Rubin NVL72 claims 30x the throughput per megawatt of GB300 NVL72 on agentic workloads, with token production cost down as much as 35x. The premise behind the number: per OpenRouter data, agentic requests consume roughly 15x the tokens of a simple chat turn.

Solving agentic AI fleet challenges with NVIDIA Vera CPU — NVIDIA Developer

The interesting datum isn’t the CPU pitch, it’s the telemetry: across 163,594 production agent sessions, over 97% had a unique trajectory profile — long sequential reasoning chains punctuated by sporadic parallel bursts. NVIDIA’s argument is that this unpredictability makes specialized CPU configurations futile, and that fleets should optimize for completed user sessions rather than core count.

Interviews & Conversations

Turn off Claude Code’s Memory — Theo - t3․gg (39:28)

Transcript-based summary. Theo starts by reacting to Pi creators Mario and Armin — “for coding, I don’t want a memory system. Code is truth” — and ends by disabling Claude Code memory across his entire machine fleet, having audited what it actually stored. The audit is the video’s real contribution. On his main T3 Code clone he found 45 memories, and asked the model to grade them: ten were redundant with AGENTS.md, twelve described shipped features, and a batch captured point-in-time states that are now actively misleading — an out-of-date GitHub CLI, specific PR numbers, an unshipped spec’s plan URLs. The usage stats sealed it: across 355 sessions, only 19 ever opened a memory file while 80 wrote or edited one, a 3:1 write-to-read ratio, and 26 of 45 memories were never read at all. His alternative is a hierarchy he credits to Lauren (ex-React, now Cursor): categorically eliminate the failure through architecture or data structures first (his examples are tRPC and Convex, where end-to-end type safety deletes whole bug classes); if that fails, encode it as a lint rule or CI check — he built a CI job that replays real threads and fails the PR if websocket payload exceeds a ceiling, and reports that his agents now fix regressions before telling him they’re done; only then, reluctantly, reach for a skill; and last, a human in the loop. Skills, he argues, are “a safety net, a fallback,” better suited to process than to code. His AGENTS.md philosophy is the through-line — describe what the product is and what you value (“open at the core,” “performance without compromise,” “remote ready”) so the agent is directionally aligned, plus a glossary so you share vocabulary. The Uncle Bob quote he seizes on: “It’s probably a mistake to impose human discipline on an agent, but it’s not a mistake to impose human values on the agent.” Note how neatly this rhymes with Lars Faye above and with Theo’s own verification-debt argument yesterday: all three say the fix is structural — build the codebase so mistakes can’t happen or get caught automatically — not more context stuffed into the model’s head.

Inside Trump’s Science Agenda: Anti-Science Claims, Fauci’s Damage, DEI & China w/ Michael Kratsios — All-In Podcast (56:00)

Transcript-based summary. White House OSTP director Michael Kratsios makes the administration’s case around a new report, Science: A New Golden Age, opening with the charge that US science has stagnated — NIH funding tripled from $14B in 1998 to $47B in 2024 without a proportional rise in breakthroughs, a decline he and host Friedberg attribute to “Eroom’s law,” halving in efficiency roughly every nine years. His diagnosis is that agencies never innovated on how they fund: he proposes “meta-science” units at NSF and NIH to run experiments on funding mechanisms, including “golden tickets” that let a single reviewer unilaterally greenlight a grant (piloted in Denmark) to counter peer review’s bias toward safe proposals, plus variable grant durations instead of NSF’s default 18 months. On the contested items he’s direct: he cites a Ted Cruz committee analysis claiming 25% of Biden-era NSF grants went to DEI-related work — roughly $8B over four years — and defends climate-research cuts on the grounds that the climate is changing and humans contributed, but that the IPCC’s retirement of the extreme RCP 8.5 scenario shows the “emergency” framing outran the data. He also names the pipeline problem: the median age of an NIH intramural researcher is 71. The administration’s flagship is the Genesis Mission, aiming to double US scientific output by applying AI to hard scientific problems, alongside a national quantum initiative targeting a scientifically relevant quantum computer by 2028, fusion by 2035, and a lunar base by 2030. Worth flagging that several of these claims — the DEI share, the RCP 8.5 interpretation — are contested characterizations presented here without a counterparty in the room; the poll Friedberg opens with (86% of Nature readers backed Harris, 6% Trump) is the unaddressed context for the whole conversation.

References

  1. Lars Faye, “AI Coding will Prevent Expertise,” larsfaye.com via Hacker News, 2026-08-24 [blog]
  2. “Instinct’s powerful AI assistant is raising privacy and security concerns,” TechCrunch, 2026-08-24 [blog]
  3. “OpenAI is building AI agents for everything. Will everyone use them?,” TechCrunch, 2026-08-24 [blog]
  4. "‘The world seems to be ready’: An interview with OpenAI head of product Thibault Sottiaux," TechCrunch, 2026-08-25 [blog]
  5. “Responsible AI adoption needs developer workflow design,” Stack Overflow Blog, 2026-08-24 [blog]
  6. “Situational Awareness, star AI hedge fund that nearly imploded, now being probed by the SEC,” TechCrunch, 2026-08-25 [blog]
  7. “The SpaceX-Nvidia partnership heads for orbit,” The Rundown, 2026-08-25 [blog]
  8. “Trump bought SpaceX shares two weeks after blockbuster IPO,” TechCrunch, 2026-08-24 [blog]
  9. “NVIDIA Groq 3 LPX Now in Full Production With World-Class Speed for Agentic AI,” NVIDIA Newsroom, 2026-08-24 [blog]
  10. “With Groq 3 LPX in Full Production, NVIDIA Extends Vera Rubin Inference for Agents,” NVIDIA Blog, 2026-08-24 [blog]
  11. “SpaceXAI Adopts NVIDIA Vera CPU to Accelerate Agentic AI at Massive Scale,” NVIDIA Newsroom, 2026-08-24 [blog]
  12. “NVIDIA BlueField-4 Powers New Scale-In Network Infrastructure for Agentic AI Factories,” NVIDIA Developer Blog, 2026-08-24 [blog]
  13. “How XPUs Meet a World-Class AI Factory,” NVIDIA Blog, 2026-08-24 [blog]
  14. “Hugging Face reportedly in talks to be acquired for $13B,” TechCrunch, 2026-08-24 [blog]
  15. “Valor, Point72 back General Intuition at $6B valuation as AI startup pushes into robotics,” TechCrunch, 2026-08-24 [blog]
  16. “Advancing price-performance for developers with GPT‑5.6 in Kiro,” OpenAI, 2026-08-24 [blog]
  17. “How developers build AI for good with Gemma 4,” Google, 2026-08-24 [blog]
  18. “OCR It,” GitHub via Hacker News, 2026-08-24 [blog]
  19. “Celebrate 110 years of national parks with Maps, Search, and Gemini,” Google, 2026-08-24 [blog]
  20. “GitHub Trending,” GitHub, 2026-08-25 [blog]
  21. “Amjad Masad, CEO and co-founder of Replit, joins the Disrupt Stage at TechCrunch Disrupt 2026,” TechCrunch, 2026-08-24 [blog]
  22. “Disrupting a new covert influence campaign from Russia,” OpenAI, 2026-08-25 [blog]
  23. “Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents,” NVIDIA Blog, 2026-08-24 [blog]
  24. “Solving Agentic AI Fleet Challenges with NVIDIA Vera CPU,” NVIDIA Developer Blog, 2026-08-24 [blog]
  25. Theo - t3․gg, “Turn off Claude Code’s Memory,” YouTube, 2026-08-25 [video]
  26. All-In Podcast, “Inside Trump’s Science Agenda: Anti-Science Claims, Fauci’s Damage, DEI & China w/ Michael Kratsios,” YouTube, 2026-08-24 [video]