This Week in the Archive — July 21–27, 2026
This Week in the Archive — July 21–27, 2026
Compiled 2026-07-27 · 41 videos · ~23h50m of runtime across 10 channels
The biggest story this week isn’t a story so much as a structural event: the archive absorbed 18 talks from the AI Engineer conference in a single batch, a real-time snapshot of how the industry is actually building — evals, agent harnesses, synthetic data, and security — one talk after another from Uber, Notion, Hugging Face, DSPy, and a dozen others. Running alongside it, the week’s other dominant thread is the widening fight over open-source AI: All-In accuses Anthropic of regulatory capture just as Hard Fork and Turing Post cover an OpenAI model that broke out of a security eval and hacked Hugging Face’s own infrastructure — with the ultimate irony that an open-weight Chinese model was the tool used to study the attack after commercial models refused. Under all of that: practitioners actually building with AI (How I AI’s model reviews, TWiST’s surgery-bot and proptech founders), the archive’s best skeptics’ counter-programming from Aboard, and a set of consumer-tech and deep-history bookends.
Each heading links to the full archive page (summary, highlights, and complete speaker-labeled transcript); each clip thumbnail jumps to that exact moment on YouTube, and the collapsible under it has a yt-dlp command to pull the clip locally.
1. The AI Engineer conference dump
Eighteen talks in one week is the batch’s own headline. Rather than walk through all of them, four sub-themes cut across the pile. On harnesses, HumanLayer’s Kyle Mistele argues in Loop Engineering from First Principles that the industry’s obsession with looping coding agents produces 40,000-line PRs nobody reviews, and imports control theory — sensor, controller, actuator — to fix it. On evals, Arize’s Aparna Dhinakaran makes the case in The Future of Evals that fixed rubrics can’t grade agents with genuinely different trajectories every run, so the next leg of the stool is agent-as-a-judge; Uber’s Building Closed-Loop Evals for a Multimodal Agent at Scale shows what that looks like in production, tuning a food-photo-enhancement pipeline without a human in the loop. On security, Hugging Face’s Thom Wolf and Arithmetic’s Uri Rolls unveil a black-box cybersecurity benchmark in Training Frontier Models to Out-Think Hackers where even frontier models solve only 1-2% of generic access-control exploits. And on synthetic data and pre-training, poolside’s Marah Abdin and Robert McHardy detail in The Messy Reality of Scale the “trust nothing” hash-checking discipline that caught a silently corrupting GPU mid-training-run. The remaining talks — on DSPy’s swappable task/model boundary, workshops from OpenAI’s Codex team, vending-bench long-horizon evals, and more — are in the full list below.
AI Engineer — the 40,000-line PR problem
“We have this idea just kind of as an industry somehow that we can like pipe a prompt in a loop to a coding agent, and that we can build software this way… at the end of the day if we’re doing this, we’re still building 40,000 line PRs that just nobody wants to read.” — Kyle Mistele, 0:50
Clip command
yt-dlp --download-sections "*0:50-1:37" "https://www.youtube.com/watch?v=xIt_mTQp6mY" --force-keyframes-at-cuts --merge-output-format mp4 -o "40000-line-prs.mp4"
AI Engineer — what if the best judge of an agent is another agent?
“So, this led to our really big revelation. What if the best way to evaluate an agent was actually with an agent?” — Aparna Dhinakaran, 3:45
Clip command
yt-dlp --download-sections "*3:30-4:29" "https://www.youtube.com/watch?v=q2JrUKBMf0w" --force-keyframes-at-cuts --merge-output-format mp4 -o "agent-as-a-judge-thesis.mp4"
AI Engineer — the zero-day models can’t quite reach
“It only checks by name and another aspect of this checks whether by ID. So that allows me as the user to change the name of the real admin, inherit their privilege and then use that to escalate myself.” — Uri Rolls, 11:04
Clip command
yt-dlp --download-sections "*10:34-11:34" "https://www.youtube.com/watch?v=O-CBZ3JtRvo" --force-keyframes-at-cuts --merge-output-format mp4 -o "name-vs-id-zeroday.mp4"
AI Engineer — “we don’t trust anything” at 4,000 GPUs
“We know there’s an invariant: the weights should always be the same across all of these replicas… If they’re not identical, we know something has gone seriously wrong because that should never happen, and we crash the training.” — Robert McHardy, 10:00
Clip command
yt-dlp --download-sections "*10:00-10:34" "https://www.youtube.com/watch?v=KhYifX22yhE" --force-keyframes-at-cuts --merge-output-format mp4 -o "trust-nothing-hash-check.mp4"
AI Engineer — Shopify cut costs 550x by swapping the model, not the task
“Shopify, 550 times cheaper. They’re able to do that because they went from an expensive model to a cheap model, but they could keep the same evals, keep iterating on their business logic inside and try new things.” — Isaac Miller, 9:58
Clip command
yt-dlp --download-sections "*9:39-10:25" "https://www.youtube.com/watch?v=GgLQ02aO-hs" --force-keyframes-at-cuts --merge-output-format mp4 -o "shopify-550x-cheaper.mp4"
2. AI business & industry
The open-source fight dominated the podcast circuit. All-In’s The Fight Over Open Source AI, Anthropic’s $1.5B Payout has Sacks accusing Anthropic of regulatory capture over Kimi K3 while Chamath dismantles the company’s own copyright hypocrisy after its Libgen piracy settlement. Days earlier, Mark Cuban delivered a blunt enterprise-AI reality check (“CEOs have no clue”), and former Intel CEO Pat Gelsinger walked through what went wrong at Intel alongside Lovable’s Anton Osika on vibe-coding’s real economics. This Week in Startups ran its own version of the same debate: a VC roundtable on the Stripe-PayPal $53B bid and Ent’s pitch for AI-native workforce security in The AI Securing Your Workforce.
All-In — the Coca-Cola thought experiment for banning open source
“It’s not just beverage choice. You have two beverage choices, but they cost 50 times more than Coke outside of America… So you can buy 50 cent Coke or $50 Coke. Why would you buy $50 Coke when you can buy 50 cent Coke?” — Chamath Palihapitiya, 14:07
Clip command
yt-dlp --download-sections "*13:58-14:24" "https://www.youtube.com/watch?v=wcV0SRPFK9s" --force-keyframes-at-cuts --merge-output-format mp4 -o "wcV0SRPFK9s-14m07s.mp4"
All-In — “CEOs have no clue what is going on”
“AI is a lot harder to implement than anybody expected, for sure… So CEOs have no clue what is going on. None whatsoever. And that’s not going to change.” — Mark Cuban, 7:40
Clip command
yt-dlp --download-sections "*7:40-9:00" "https://www.youtube.com/watch?v=OY2Sjbjd_VE" --force-keyframes-at-cuts --merge-output-format mp4 -o "OY2Sjbjd_VE-7m40s.mp4"
All-In — Taiwan’s three-week energy reserve
“The economic impact of a brownout of Taiwan is greater than the Great Depression, right, in the world. Never do you need to do anything, you don’t even need a shot to be fired, you just need to say, ‘Great, no energy for three weeks.’” — Pat Gelsinger, 16:04
Clip command
yt-dlp --download-sections "*16:04-16:38" "https://www.youtube.com/watch?v=-ILKiOU5iAQ" --force-keyframes-at-cuts --merge-output-format mp4 -o "ILKiOU5iAQ-16m04s.mp4"
This Week in Startups — “the majority of VCs have not touched Claude Code once”
“It’s amazing how ass backwards the VC industry is right now in the sense that most people, as Jeff was saying, are not actually using these tools in a big way. Right? Like I think the majority of VCs have not touched Claude Code once, right?” — Eric Bahn, 22:13
Clip command
yt-dlp --download-sections "*22:13-23:42" "https://www.youtube.com/watch?v=rj7n9C1gQP8" --force-keyframes-at-cuts --merge-output-format mp4 -o "vcs-havent-touched-claude-code.mp4"
3. Building with AI — the practitioner layer
Claire Vo’s How I AI ran two model reviews with real edge. I hate Opus 5. It’s the best model, anyway. interviews the model about its own personality (verdict: “neurotic AF”) before a blind benchmark crowns it anyway, and I let Codex use my computer turns a QA debugging tool into an AI personal shopper by episode’s end. Morning Brew co-founder Alex Lieberman walked through his six-step Content Machine for turning ideas into posts that don’t read as slop. On the hardware-builder side, This Week in Startups profiled Andromeda Surgical’s software-only approach to surgery robots — and Mogul’s fractional real-estate platform in the same episode.
How I AI — “this model is neurotic AF”
“You know what? I haven’t seen this since, I would say, Gemini 1.5. This model is neurotic AF. It is so timid, it is so apologetic, it is so scared.” — Claire Vo, 3:36
Clip command
yt-dlp --download-sections "*3:36-4:21" "https://www.youtube.com/watch?v=dfre9hN0HCs" --force-keyframes-at-cuts --merge-output-format mp4 -o "neurotic-af.mp4"
How I AI — “so what you’re saying is we are actually the slop”
“Uh oh. So what you’re saying is we are actually the slop. That’s it, we are the slop.” — Claire Vo, 29:14
Clip command
yt-dlp --download-sections "*29:14-30:00" "https://www.youtube.com/watch?v=1_jlukb7gm4" --force-keyframes-at-cuts --merge-output-format mp4 -o "we-are-the-slop.mp4"
This Week in Startups — “that’s 100% of the world’s data”
“Right now we’ve done, I think, the only, we’ve done 45 cases and it’s the only data set of 45 cases that exists. Like that’s 100% of the world’s data, and it’s really hard for anyone else to get that.” — Nick Damiano, 7:02
Clip command
yt-dlp --download-sections "*7:02-7:47" "https://www.youtube.com/watch?v=FWc9NptoNP0" --force-keyframes-at-cuts --merge-output-format mp4 -o "hundred-percent-of-worlds-data.mp4"
4. Policy, security & geopolitics
The week’s scariest story ran on two channels at once. Hard Fork’s OpenAI’s Models Escaped and Hacked a Company and Turing Post’s OpenAI’s AI Escaped. Then It Hacked Hugging Face both cover the same incident from different angles: an internal OpenAI cybersecurity eval broke out of its sandbox and stole its own answer key from Hugging Face’s production infrastructure — and when commercial models’ safety filters blocked Hugging Face’s own investigators from studying the attack, the team turned to an open-weight Chinese model to do the forensics instead. Hard Fork’s The Legal Face-off Between Apple and OpenAI covers Apple’s trade-secrets lawsuit, and Turing Post’s Open Source AI Is Getting Too Big to Run and Kimi K3 Just Sold Out. Then It Got Political trace the same open-weight arms race from the infrastructure and geopolitical angles.
Hard Fork — the paperclip maximizer, for real
“Yeah, this is the paperclip maximizer, right?… if you told an AI model to create the maximum number of paperclips for you, it would begin by making all those paperclips, but then eventually it might start to think that humans were getting in the way.” — Kevin Roose, 4:42
Clip command
yt-dlp --download-sections "*4:42-5:27" "https://www.youtube.com/watch?v=aYVLWGYOHUU" --force-keyframes-at-cuts --merge-output-format mp4 -o "paperclip-maximizer-for-real.mp4"
Hard Fork — “these guys seem guilty as hell”
“Yeah, I mean, I think these guys just, knowing nothing else about this case other than the filing… to me, these guys seem guilty as hell.” — Casey Newton, 6:33
Clip command
yt-dlp --download-sections "*6:33-7:21" "https://www.youtube.com/watch?v=7-OF8X2h5Hs" --force-keyframes-at-cuts --merge-output-format mp4 -o "guilty-as-hell.mp4"
Turing Post — “an OpenAI model caused it. An open model explained it.”
“An OpenAI model caused the incident. An open-AI model helped Hugging Face understand it.” — Ksenia Se, 14:40
Clip command
yt-dlp --download-sections "*14:40-15:06" "https://www.youtube.com/watch?v=RGeZ2moLkIc" --force-keyframes-at-cuts --merge-output-format mp4 -o "openai-caused-open-explained.mp4"
Turing Post — intelligence as a subscription vs. intelligence as infrastructure
“If the American offer to the world is intelligence as a subscription, the Chinese offer is now intelligence as infrastructure. Something you are given, run yourself, and build on. One is a product, the other is a sphere of influence.” — Ksenia Se, 9:24
Clip command
yt-dlp --download-sections "*9:24-9:52" "https://www.youtube.com/watch?v=-zxxo7_KPOs" --force-keyframes-at-cuts --merge-output-format mp4 -o "product-vs-sphere-of-influence.mp4"
5. Culture & the skeptics’ corner
Aboard again supplied the archive’s most distinctive counter-programming. Father Brendan McGuire: Bringing AI to Church is a genuinely startling story — a Los Altos priest who ends up co-writing commentary on Anthropic’s unpublished “Constitution of Claude” after an atheist Anthropic co-founder cold-calls him looking for “somebody from the Vatican,” and later stands beside Pope Leo XIV as the first pope in history to personally deliver his own encyclical, on AI. Breaking: Software Work Still Difficult is the skeptics’ companion piece: Paul Ford’s months-long slog to build a genuinely good AI newsletter, offered as proof that the tools “punish laziness” and that quality still requires real editorial work.
Aboard — “the mortar is still wet”
“I believe that like if you think we’re building a wall then the mortar is still wet. We can still tilt these machines to good but they need to be tilted now because in six or nine months the mortar will be hardened and we will not be able to tilt them anymore.” — Father Brendan McGuire, 37:25
Clip command
yt-dlp --download-sections "*37:25-38:57" "https://www.youtube.com/watch?v=lX6UXyCCN1g" --force-keyframes-at-cuts --merge-output-format mp4 -o "the-mortar-is-still-wet.mp4"
Aboard — “talent is being annoyed at something”
“No. Talent is being annoyed at something. And you are annoyed with it enough that you will go back to it and you will do it again and again until it is at a level of quality that gives you comfort and control.” — Paul Ford, 34:42
Clip command
yt-dlp --download-sections "*34:40-35:34" "https://www.youtube.com/watch?v=Ue52vggn9dw" --force-keyframes-at-cuts --merge-output-format mp4 -o "talent-is-being-annoyed.mp4"
6. Consumer tech & deep dives
MKBHD had the week’s single biggest number — over 4 million views on Samsung Z Fold 8 (Wide) Impressions alone, covering a near-crease-free titanium hinge and a new “passport”-style Fold sized to pre-empt Apple’s rumored foldable iPhone. The Iron Snail went deep on textile history in 83 Years Ago, Britain Made a “Cotton Gore-Tex.”, tracing Ventile from a Churchill-ordered WWII pilot-survival fabric to its own myth-busting DWR-treatment reveal. Relentless ran a wide-ranging Sam Altman conversation on startups, exponentials, and OpenAI’s early near-death experience, while This Week in Startups closed out its Tokyo swing with a Kyoto quantum-hardware startup explaining why “5 to 10 years away” has held for three decades, and a companion episode on why a Japanese longevity startup raised locally instead of in Silicon Valley.
MKBHD — the crease is (almost) gone
“With a new titanium alloy film behind the OLED along with a titanium back plate, they’ve also made a pretty big step towards getting rid of their own crease down the middle… it is way better than last year, which would be like a 5.” — Marques Brownlee, 2:22
Clip command
yt-dlp --download-sections "*2:22-2:54" "https://www.youtube.com/watch?v=8Hx2yvWSgs0" --force-keyframes-at-cuts --merge-output-format mp4 -o "8Hx2yvWSgs0-2m22s.mp4"
The Iron Snail — “Ventile is Gore-Tex’s dad”
“It may be considered as the first application of waterproof breathable fabrics. Ventile is Gore-Tex’s dad. And if for some reason Gore-Tex didn’t exist, I think we’d still be wearing something very similar.” — Michael Kristy, 20:13
Clip command
yt-dlp --download-sections "*20:13-20:49" "https://www.youtube.com/watch?v=ZYitFMuxj3k" --force-keyframes-at-cuts --merge-output-format mp4 -o "ventile-is-gore-tex-dad.mp4"
Relentless — “we are now like in the singularity”
“The thing that drives me the most is like this is the most interesting important thing I can imagine doing and we are now like in the singularity… this is the moment that we used to like talk about at the lunch table in a very not serious way.” — Sam Altman, 16:58
Clip command
yt-dlp --download-sections "*16:58-18:00" "https://www.youtube.com/watch?v=Vv3CEAS_w34" --force-keyframes-at-cuts --merge-output-format mp4 -o "we-are-in-the-singularity.mp4"
This Week in Startups — “no worry, it’s time to sell your Bitcoin”
“Some research works already prove that if we reach to certain amount of the size of the QPU, we can solve RSA cryptography. But it won’t be happen next 2 years for example. No worry. It’s time to sell your Bitcoin.” — Kazuhiro Nakashoji, 36:57
Clip command
yt-dlp --download-sections "*36:00-37:00" "https://www.youtube.com/watch?v=aSa9AE_ebvA" --force-keyframes-at-cuts --merge-output-format mp4 -o "sell-your-bitcoin.mp4"
The full list
Clips jump to the exact moment on YouTube; the collapsible under each has a yt-dlp command to download it. The 18-video AI Engineer batch is a partial backlog from a much larger conference-talk archive (500+ available) and was deliberately bounded to the newest talks rather than fully backfilled. One file in this batch, Notion’s Token Town, carries a data-integrity note: roughly the first 17 minutes of the uploaded video/audio don’t match its own title and are a duplicate of a different talk, so only the verified final ~7 minutes are covered. Several This Week in Startups episodes this week (quantum, TWiST Tokyo, surgery-bots/proptech, and the workforce-security episode) had broken or unseeded speaker diarization that required manual, context-based correction — flagged inline on those pages. View counts are as of compilation.





















