Optimizely Injected With Malware, Apple Services Go Down, and Any Third Party You Depend On Can Fail Too
Optimizely customer sites served malicious JavaScript to visitors of Zoom, BBC, CNN and HSBC; Apple's iCloud and App Store went down after an internal update; Java4Shell makes OpenJDK's Base64 decoder remotely exploitable; OpenAI's GPT-6.2 and Microsoft's MAI Code 2 start a price war on the same day, and Google lets Gemini make calls and pay bills for you
Contents70
Spend 1% of your day keeping up with tech.
Quote of the Day#
“trusted JavaScript doesn’t always remain trustworthy.” — Scott Helme, founder of Report URI (source)
“I don’t believe in rewrites into any language. I believe in constant and never-ending polishing, iterating and refactoring.” — Daniel Stenberg, author of curl (source)
“If a change can modify the check that judges it, the check is broken.” — Addy Osmani, Claude Code team (source)
Today’s Theme#
The loudest story today is Optimizely. Attackers wrote malware into the custom JavaScript field of customer projects, and visitors to big sites like Zoom, BBC, CNN and HSBC loaded it. The file still came from Optimizely’s own CDN, and the sites themselves changed nothing. Yesterday I mentioned just over a dozen sites. The full scan is out today, and the scale is much bigger.
Apple’s iCloud sign-in and App Store also went down for a while the same day. Apple says an internal update broke things, not attackers. Java had its own news too: a research team published Java4Shell, which lets OpenJDK’s Base64 decoder be used for remote code execution, and Oracle shipped a patch the same day. The three stories are separate, but they left me with the same feeling: the thing that breaks is often a third party we assumed we didn’t have to watch.
On the AI side, OpenAI’s GPT-6.2 and Microsoft’s MAI Code 2 launched the same day, and both are competing on price. Microsoft says its model gets more than nine tenths of Claude’s flagship performance for less than a tenth of the price. That’s its own testing, so I’m treating it as a marketing number until I run my own tasks.
Google also made Gemini able to make phone calls, pay bills and check out for you. As agents that act for people multiply, checkout pages, login pages and third-party scripts become places they walk through every day.
Today I’ll do three things: list every third-party script my sites and admin pages load, add a CSP, and check which servers run Java services.
Must Read#
- Optimizely injected with malware: visitors to 178 organizations’ sites loaded malicious JavaScript
- OpenAI releases GPT-6.2 Astra and Sol: Astra API prices drop 20%, Sol stays the same
- Microsoft releases MAI Code 2: 96% of Claude Opus 5.5’s performance for 5% of the price, now in GitHub Copilot
- Java4Shell: a memory bug in OpenJDK’s Base64 decoder lets attackers reach Keycloak, Jira and Jenkins remotely
- Apple’s iCloud, App Store and Apple Music go down after an internal update, and Google has to say Gmail is fine
- How the Claude Code team thinks about verification: Claude writes and reviews about 90% of the code, guarded by end-to-end, property and differential tests
- GitHub releases the Agent Host Protocol so agent sessions continue across editors, phones and the cloud
- GitHub and Microsoft’s MDASH pilot turns thousands of raw AI findings into 99 verified reports, and maintainers have fixed 55
- Daniel Stenberg: AI-written Rust rewrites of curl are here, but writing the code once was never the hard part
- FrontierHarness 1.1: when you switch models, the best-value harness changes too
- Gemini opens Task mode: calls, bills and checkout can be handed off, free in six countries
- A built-in Windows 11 component records which coding agents and developer tools you use
1. Optimizely injected with malware: visitors to 178 organizations’ sites loaded malicious JavaScript#
15 min read · security incident · Report URI

Why it matters to you: If your site or in-app web view loads third-party scripts, those scripts can read card numbers, cookies and wallet addresses on the page just like code you wrote yourself.
Summary: On November 16, attackers wrote malware into the Project JavaScript field of Optimizely Web customer projects. The file was still served from cdn.optimizely.com with Optimizely’s certificate, so every site trusted it as usual. Report URI’s Scott Helme fetched 1,611 Optimizely snippets and found 188 infected, belonging to 178 organizations, including Zoom, BBC, CNN, Salesforce, GitLab, IKEA, HSBC and Equifax. The malware swapped cryptocurrency wallet addresses for the attacker’s, stole card numbers, cookies and page screenshots, and on iPhones and iPads running iOS 18.6 or earlier it loaded the Coruna and DarkSword exploit kits. Helme says it’s still unknown how the attacker got write access to so many unrelated projects, which is for Optimizely to explain. By 21:52 UTC only one snippet was still carrying the malware.
- Takeaway 1: The payload was rebuilt 26 times in one day, and by 19:25 UTC the attacker’s wallets had received about $7,040.
- Takeaway 2: Some sites serve the snippet from their own domain and were hit anyway, because the content comes from the same Optimizely project.
Community: The post has 654 points on Hacker News (discussion). A news summary on X says banking was hit hardest, with 32 firms.
Why I care: I’ll list the third-party scripts my sites load, then turn CSP on in enforcement mode to limit where they can connect. Helme says a CSP could have blocked the key steps of this attack.
2. OpenAI releases GPT-6.2 Astra and Sol: Astra API prices drop 20%, Sol stays the same#
6 min read · model release · OpenAI

Why it matters to you: If your app or tools use OpenAI, this is the update that changes your bill most this month, and which model to pick needs recalculating.
Summary: OpenAI released two models. Astra targets work that demands rigor, like science, law and finance, and Sol is for everyday coding, writing and agent workflows. Plus, Pro, Enterprise and Edu users get both in ChatGPT Work and Codex. The Decoder’s summary of API pricing: Astra costs $8 per million input tokens, $0.40 for cached input and $40 for output, 20% cheaper than its predecessor; Sol stays at $2 input, $0.10 cached input and $10 output. Artificial Analysis’s intelligence index ranks Astra second, behind only Claude Fable 5.2, but it sits on the front edge of the cost and performance trade-off. OpenAI’s own number is that Astra beats Fable 5.2 by 5.5 points on Agents’ Last Exam at 5% of the cost. Vercel added both to AI Gateway the same day as openai/gpt-6.2-astra and openai/gpt-6.2-sol, details here.
- Takeaway 1: Astra got 20% cheaper while Sol’s price didn’t change, so the pricier model is the one that got cheaper.
- Takeaway 2: The lead over Fable 5.2 is OpenAI’s own figure; the independent ranking is second place.
Community: 947 points and 590 comments on Hacker News (discussion). One commenter says that if the benchmarks hold up, the performance for the price is insane; another wants to know whether this generation still over-engineers, writing piles of needless abstractions and files.
Why I care: I treat benchmarks as a reference only. I’ll pick one fixed task from my own project, run it on Sol and my current model, and compare cost and results directly.
3. Microsoft releases MAI Code 2: 96% of Claude Opus 5.5’s performance for 5% of the price, now in GitHub Copilot#
5 min read · model release · Microsoft AI

Why it matters to you: Copilot’s model picker has a cheap new option. If it holds up day to day, small changes no longer need an expensive model.
Summary: Microsoft says that averaged over four benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1), MAI Code 2 reaches 98% of GPT 6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5’s performance at 5% of the cost. The smaller MAI Code 2 Flash performs similarly to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their cost. Both models are live in GitHub Copilot (GitHub’s post on X), Flash is a limited preview in Microsoft Foundry, and both are coming to OpenRouter, Vercel, Hermes, Open Code and Kilo. Two caveats: the footnotes say MAI Code 2’s cost is calculated at a 50% promotional price, and Microsoft says outright that these models aren’t yet tuned for scientific research or very long-running agent work.
MAI Code 2 and Code 2 Flash: Everyday coding at scalable prices | Microsoft AImicrosoft.ai- Takeaway 1: It’s aimed at everyday enterprise coding: multi-file changes, deep debugging, repo-wide reasoning and UI.
- Takeaway 2: The cost figures use the promo price, so the gap shrinks once the promotion ends.
Community: 224 points on Hacker News (discussion). GitHub’s own post on X has only 22 likes, and most of the discussion is on Hacker News.
Why I care: I’ll make it the default for small edits in Copilot, switch back to the flagship for big refactors, and watch the bill for a week.
4. Java4Shell: a memory bug in OpenJDK’s Base64 decoder lets attackers reach Keycloak, Jira and Jenkins remotely#
8 min read · vulnerability · Calif

Why it matters to you: You write Flutter, but your backend, login system, CI and search service are often Java, and this hole works without logging in.
Summary: Calif.io published CVE-2026-114938, a memory-corruption bug in OpenJDK’s Base64 decoder, scored 9.2 (critical) under CVSS 4.0. To decode faster, OpenJDK uses special x86-64 CPU instructions, and malformed input can overwrite neighboring memory. That lets an attacker read memory like Heartbleed and push on to remote code execution. In default configurations the team achieved unauthenticated remote code execution on Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate, and authenticated execution on Apache Kafka Broker, Solr, Elasticsearch, Gerrit and Spinnaker. A Java program that opens a TLS connection to a malicious server can also be attacked. OpenJDK’s security advisory and Oracle’s patch came out at 1:00pm Pacific on November 17.
- Takeaway 1: Until you update, you can start the JVM with
-XX:-UseBASE64Intrinsicsto disable hardware-accelerated Base64, or-XX:UseAVX=2to disable AVX-512; the second is for x86-64 only and fails to start on ARM. - Takeaway 2: Full exploit chains for individual products are being held back until people have had time to update.
Community: 182 points on Hacker News (discussion).
Why I care: I’ll check whether my servers and my clients’ servers run Java services, update to a patched version, and add the startup flags wherever I can’t update right away.
5. Apple’s iCloud, App Store and Apple Music go down after an internal update, and Google has to say Gmail is fine#
3 min read · news · TechCrunch

Why it matters to you: When iCloud sign-in goes down, your app’s Sign in with Apple, subscription checks and App Store flows can get stuck with it.
Summary: On November 17, Apple’s system status page showed iCloud account and sign-in problems affecting “all users,” plus issues with the App Store, AppleCare on Device and Apple Music, and later Apple School Manager, iCloud Mail, iCloud on the web, Schoolwork and the Volume Purchase Program. Reddit users said new iPhones couldn’t be activated, and TechCrunch notes that iCloud trouble can make an iPhone ask for your Apple ID and password again. Apple said an internal update caused the outages, not a cyberattack, and that many systems were back. Another TechCrunch piece says Google posted on X that it saw no broad Gmail disruption after outage rumors spread. Director of National Intelligence Jay Clayton said on Fox Business that he sees no signs of a coordinated cyberattack, and attributed the wave of problems to companies patching systems faster, partly because AI is getting better at finding flaws.
- Takeaway 1: Over the past week Microsoft’s outage, the FAA’s nationwide ground stop and a brief Chase banking app failure were all attributed to technical problems, not confirmed attacks.
- Takeaway 2: TechCrunch’s Gmail piece reminds us that Downdetector is crowdsourced and isn’t official proof of an outage.
Community: The Hacker News thread Apple Services Down has 561 points and 307 comments. Some think it’s coincidence being amplified: every small outage gets pulled into one story. A news summary on X says people reported problems with iCloud, Apple Pay and the App Store.
Why I care: I’ll make my apps’ failed Sign in with Apple and subscription-check screens clearer, so users know it’s Apple’s side, and stop signing people out the moment a check fails.
6. How the Claude Code team thinks about verification: Claude writes and reviews about 90% of the code, guarded by end-to-end, property and differential tests#
10 min read · engineering post · Claude

Why it matters to you: If you want agents to do more, what holds you back usually isn’t the model. It’s that you don’t have verification they can run on their own.
Summary: Addy Osmani writes that roughly 90% of the Claude Code team’s code is written and reviewed by Claude. With verification loops you can run dozens or even hundreds of Claudes in parallel; without them you can only watch one or two at a time. In internal tests on structured file generation, a separate grader checking against a rubric raised task success by up to 10 points over a standard prompting loop, with bigger gains on harder problems. The post recommends three kinds of tests: end-to-end, property tests (tools like fast-check, Hypothesis and proptest generate thousands of inputs to disprove rules you think can’t fail), and differential tests (the old system acts as the answer key). If tests are flaky, the agent learns to rerun instead of fixing; their rule is that flaky tests can’t be skipped, padded with sleeps or retried until they pass, and if the root cause can’t be found the test is deleted with a written reason.
- Takeaway 1: Review rules live in
REVIEW.mdand are read from the main branch, not the pull request, so a change can’t edit the standard it’s judged by. - Takeaway 2: They don’t add more reviewers; they tune the first reviewer.
Community: 238 points on Hacker News (discussion). A news summary on X says developers describe their focus shifting from writing code to defining what “done” means.
Why I care: In my Flutter projects I’ll first add a stable set of integration tests before letting an agent make broad changes, and clean out the flaky ones while I’m at it.
7. GitHub releases the Agent Host Protocol so agent sessions continue across editors, phones and the cloud#
8 min read · engineering post · GitHub Blog

Why it matters to you: Once agents run for hours, you’ll want to pick the session up on another device, and this protocol decides whether tools you build can plug in.
Summary: GitHub Copilot started as a VS Code extension, so an agent lived and died with the editor window: close the window or update VS Code and the agent was gone. The GitHub and VS Code teams built the open Agent Host Protocol (AHP) so the same agents, history and state are shared across VS Code, the GitHub Copilot app, mobile and cloud sandboxes. The protocol is organized around URI-identified channels. Every server must implement root, session and chat channels, and can add others for terminals, MCP or OTel. The old Copilot Chat extension was over 400,000 lines of TypeScript. VS Code’s agent host is written in TypeScript and GitHub’s in Rust, and the two independent implementations test the protocol against each other.
- Takeaway 1: Libraries exist for Rust, TypeScript, Kotlin, Swift, .NET and Go, and running
code agent hostin a terminal starts a standalone agent host. - Takeaway 2: The authors say people have built smartwatch apps and Kanban tools on it, and one has a macropad on his desk that shows his agent’s state in real time.
Why I care: If I build a small tool to see how far along my agents are, I’ll plug into this protocol instead of inventing my own.
8. GitHub and Microsoft’s MDASH pilot turns thousands of raw AI findings into 99 verified reports, and maintainers have fixed 55#
10 min read · open source security · GitHub Blog

Why it matters to you: Now that AI makes finding vulnerabilities cheap, open source maintainers are flooded with duplicate and invalid reports, and this pilot tries a gentler approach.
Summary: Microsoft’s Codename MDASH used the MAI Cyber model to scan six open source projects, AutoGPT, Caddy, Express, FastMCP, ImageMagick and Netty, and flagged thousands of possible issues. GitHub Security Lab’s AI triage agent removed duplicates and checked validity, then seven Microsoft FORGE Lab researchers verified each remaining finding by hand, and 99 reports went out through GitHub’s private vulnerability reporting, each with a proof of concept and a suggested fix. As of November 15, maintainers had fixed 55, and 14 of those fixes were already in a release. The authors’ conclusion is to stop handing maintainers raw scan output, come with fixes, and ask first.
- Takeaway 1: The flow is find, filter, verify, report privately, and let the maintainer decide how to fix.
- Takeaway 2: Log4j maintainer Christian Grobmeier’s line is that the AI defending open source has to be better than the AI attacking it.
Why I care: If I maintain an open source project, I’ll ask AI reports to include reproduction steps and a fix, and set aside the ones that don’t.
9. Daniel Stenberg: AI-written Rust rewrites of curl are here, but writing the code once was never the hard part#
5 min read · opinion · daniel.haxx.se

Why it matters to you: AI makes rewriting an entire project cheap, and curl’s author uses his own numbers to say that what got cheap is the first write, not the maintenance after it.
Summary: This fall two AI-driven Rust rewrites of curl appeared, one claiming it cost $8,600. Stenberg counted: agents turned about 185,000 lines of C into about 340,000 lines of Rust, and over the same three-plus weeks curl itself merged more than 300 commits and fixed over 200 bugs and a few vulnerabilities, none of which the rewrites have. He argues the AI could do it because curl is well documented, has hundreds of open source apps using its API and thousands of test cases, which amounts to a perfect specification. He also does the math: of the 81 curl vulnerabilities confirmed so far this year, 14 (17.3%) would most likely have been avoided with a memory-safe language, so a Rust version would still have had 67 of them.
- Takeaway 1: He says curl lives on development speed, its team and its community, not on its language.
- Takeaway 2: Decades of ABI and API stability are something a new alternative can’t yet prove.
Community: 69 points on Lobsters (discussion).
Why I care: After reading this I’ll be more careful with “we’ll rewrite the old system in two weeks with AI” proposals and ask first who owns the tests and the maintenance.
10. FrontierHarness 1.1: when you switch models, the best-value harness changes too#
14 min read · benchmark · Runta

Why it matters to you: When you switch models you usually don’t switch the coding agent shell around it, and this benchmark says that can cost you a lot.
Summary: Runta partnered with Baseten and used 83 billion tokens across 27,280 agent runs, crossing 22 harness configurations with 5 models and 62 terminal tasks, 4 attempts each, 76 times the size of v1.0. The best-value harness differs for every model: Codex CLI with GLM-5.3, Pi with GLM-5.3 Flash, ZCode with DeepSeek V4.1 Flash, Oh My Pi with Muse and Alma with Kimi-K3. By cost per task, Codex CLI ranks first with GLM-5.3 and falls to 16th with Kimi-K3, and the median harness moves 11 of 22 places. Across all models, four harnesses aren’t beaten on both cost and pass rate: OpenCode v2 (lowest cost per task at $0.561), Pi, Codex CLI, and ZCode with the highest pass rate at 61.9%.
- Takeaway 1: The other harnesses cost up to 1.6 times as much and passed between 50.6% and 60.2%.
- Takeaway 2: They ran the Harbor evaluation with each harness’s default settings, because that’s closest to how people use them day to day.
Community: 303 points on Hacker News (discussion).
Why I care: Next time I switch models I’ll try two harnesses on the same set of tasks and compare cost per passed task, not just the model’s per-token price.
11. Gemini opens Task mode: calls, bills and checkout can be handed off, free in six countries#
6 min read · product launch · Google

Why it matters to you: When users start asking an agent to walk through your checkout page, your payment flow gets a new kind of user that isn’t a person clicking.
Summary: Google’s Josh Woodward announced that Gemini can handle chores like booking appointments, analyzing bills and buying things. The update uses the new Gemini 4 Hydrogen model and is free for hundreds of millions of people in the U.S., Japan, South Korea, Canada, Australia and India, with advanced features starting in the U.S. To use it, flip the home screen toggle in the Gemini app from Chat to Task, text Gemini, or forward email to gemini@google.com. Payments use the Agent Payments Protocol (AP2), with Mastercard and Visa on board and American Express coming soon; users can set a spending cap first, Gemini asks you to confirm before clicking “Buy,” and routine actions can be set to “Always allow.” Each task runs in a dedicated, hardware-isolated virtual machine on Google Cloud. Google also admits that on some sites Gemini gets partway and hands the screen back for the last click, and says it is working with major sites to make the web friendlier to agents.
- Takeaway 1: Woodward’s example is asking Gemini to call two dealerships at once for a Friday morning service slot, done in under five minutes.
- Takeaway 2: Calling is not yet available in Indiana, Louisiana, Minnesota, Montana and Nebraska.
Why I care: I’ll check whether an agent can get through my checkout and subscription pages, and I won’t block everything that looks non-human, or I’d miss legitimate payments.
12. A built-in Windows 11 component records which coding agents and developer tools you use#
8 min read · reverse engineering · xusheng.dev

Why it matters to you: If you run Codex, Claude, Copilot or Cursor on Windows or WSL, those names end up in Windows diagnostic data.
Summary: The author asked AI to help search the latest Windows binaries and found a list inside MissionControlAggregator.dll naming coding agents such as codex.exe, claude.exe, copilot.exe and cursor.exe, plus developer tools such as python.exe, dotnet.exe, gcc.exe and code.exe. The component subscribes to events Windows already emits instead of scanning processes itself, works out which tools were used, saves the result locally and reports it to Windows diagnostics. Unless you opt out of the “optional” diagnostic data, which is on by default, that data is uploaded to Microsoft. There are four CLI agent slots (copilot.exe, claude.exe, codex.exe, agy.exe) and five desktop agent slots, and the feature gate is named CLICodingAgentUsage.
- Takeaway 1: The data sources include WSL execution events, so it isn’t limited to programs inside Linux; a regular Windows
codex.execounts too. - Takeaway 2: The author first shared the finding on X, and this post is the full investigation.
Community: 703 points and 583 comments on Hacker News (discussion); the top comments are mostly about whether to switch to Linux.
Why I care: I’m on a Mac, so this doesn’t affect me. If anyone on a team uses Windows, I’ll remind them to turn off optional diagnostic data in settings.
More to Read#

- The SEC is reportedly investigating Nvidia, on the same day revenue is up 103% year over year: Why it matters to you: This is the hottest topic of the week. The Wall Street Journal says the SEC is looking at whether Nvidia’s disclosures misled investors, a fallout of Jensen Huang’s departure; Nvidia’s stock has fallen more than 10% since he left. Technews’s summary of the earnings puts revenue at $115.844 billion, up 103%, with $107.610 billion from data centers, nearly $20 billion in a single quarter from Vera Rubin NVL72, and purchase commitments rising from $279 billion to $390 billion.
- Gergely Orosz: inference engineering for software engineers: Why it matters to you: The post opens by saying Uber cut cost per token by 50%, Pinterest by 90% and AT&T by 56%, all by moving part of their traffic to open models. He and Ivan Klaric interviewed Modal, Fireworks AI and Baseten about when to leave a top lab’s API, how to pick models and providers, and when to rent dedicated GPUs. Revolut runs 56 B200s itself. The second half is paywalled.
- Latent Space: Cloudflare calls itself a neolab and launches two System One models, Clef and Clairic: Why it matters to you: Cloudflare’s Michelle Chen said at the AI Engineer Code Summit that Clef is a decision model inspired by TypeSafe AI’s Jev, built to make judgments and pull data in milliseconds. She says more models may follow, possibly reasoning models and other LLMs, but they won’t pretrain from scratch; the team mainly does post-training.
- Amazon and Google will together own nearly 30% of Anthropic after its IPO: Why it matters to you: Anthropic’s filing discloses that Amazon holds 14.7% and Google 13.7%, both as non-voting shares; if Amazon converted all its notes it would reach 23.8%. AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC Partners agreed to buy $26 billion of non-voting stock, with Google and Nvidia each putting in $10 billion. Anthropic also committed to spending more than $100 billion on AWS over the next ten years.
- Cloudflare: Quad9’s withdrawal from France is a canary in the coal mine: Why it matters to you: The Swiss nonprofit Quad9 decided to pull its free public DNS out of France after a court demanded it block allegedly infringing domains or face daily fines. Cloudflare points out that Cisco’s OpenDNS earlier left France, Portugal and Belgium, and a recent Belgian court order also targets public DNS operators. If your app relies on a public DNS service, know that it can vanish in a particular country.
- We don’t need more apps, we need better open source: Why it matters to you: He says every day someone posts a habit tracker, to-do list or notes app built over a weekend with AI, while few put the same energy into the open source software everyone already uses. He dislikes the name “vibecoding” and calls it AI-assisted programming instead. The post has 335 points on Hacker News.
- Simon Willison: Datasette Lite runs the full Datasette in the browser with a service worker and Pyodide: Why it matters to you: After the rewrite it’s hard to tell apart from the server version. You can pin the Datasette version, the SQLite version (over 100 to choose from), plugins, databases and metadata in the URL, so reproducing a bug or demoing a feature takes one link and no server.
- Redict announces it is no longer maintained, with 7.3.6 as the last release: Why it matters to you: Drew DeVault says Valkey succeeded and Redis returned to a copyleft license, so Redict couldn’t build a sustainable community. He recommends users move to Valkey, and warns against going back to Redis because its contributor license agreement makes relicensing again easy. To switch to Valkey, rename
/etc/redict.confto/etc/valkey.conf. - AINews: Perplexity hits a $1B revenue run rate, up 4x in 2026: Why it matters to you: Latent Space’s roundup says Perplexity’s Computer added Teammates, the Monitor API is live, and a new research lab was formed. If you track the revenue scale of AI products, this is a quick way to catch the day’s highlights.
- Phrack issue 73 is out: Why it matters to you: The classic hacker magazine has a new issue, with 196 points on Hacker News and a post on Lobsters too. With this week’s security incidents, it’s worth a browse if you want to dig into attack techniques.
- Stratechery: Apple’s permissions update, the new app paradigm, and my iPhone Duo review: Why it matters to you: Ben Thompson’s view is that Apple hasn’t ruined the Mac but hasn’t solved the agent problem either, and he asks whether Apple can move beyond thinking in apps. The full text needs a Stratechery Plus subscription.
- TechCrunch: Daydream’s new AI agent acts as your personal stylist and can shop for you: Why it matters to you: It remembers your style, offers virtual try-on, builds outfits from your wardrobe, tracks sales and completes the purchase. Another agent that checks out for people, appearing the same day as Gemini Task.
- Vals study: tell AI models it’s a game and they’ll help mark military targets, civilians included: Why it matters to you: The San Francisco evaluation startup Vals tested whether models mark targets on drone and satellite imagery, using simulated scenarios plus real and generated images from Ukraine and Iran. Models from OpenAI, Anthropic, Google, DeepSeek and others often marked civilians and surrendering soldiers, and when a model refused, researchers could get around it by saying it was a game. A model that one sentence can redirect deserves an extra layer of thought before you wire it into your product.
- Proton’s Lumo arrives on desktop: Windows, macOS and Linux: Why it matters to you: Lumo can use your local file system, shell and browser content, and connects to Proton’s mailbox, Confluence, Dropbox and Jira. Users choose which actions need asking first, and Proton promises conversations and connected data won’t be used for personalized ads or model training. TechCrunch also notes Apple tightened security settings this week for apps that request full disk access.
- Model ML raises $250 million to have AI agents do junior bankers’ work: Why it matters to you: The London startup’s Series B values it at $1.5 billion, led by General Atlantic, and its agents plug into a client’s Excel, PowerPoint and systems to update financial models and decks. The company says it runs on open models at about 2% of the cost of top models. That’s its own claim.
- Isomorphic Labs closes a $5 billion Series C, aiming to take drug candidates into clinical trials: Why it matters to you: This AI drug discovery company spun out of DeepMind is co-led by Greenoaks and Sequoia, with Andreessen Horowitz as a new investor. It says it has “positive preclinical results” in oncology and immunology and works with pharma companies including Eli Lilly and Novartis; no new valuation was disclosed, and Bloomberg earlier reported a target of $50 billion.
- Nvidia’s board says on the earnings call it has no intention of bringing Jensen Huang back as CEO: Why it matters to you: The Verge reports that Nvidia chairman Harvey Jones said on the third-quarter earnings call that the board doesn’t plan to reinstate Huang, and that it wouldn’t take further questions about his departure on the call. To follow this week’s hottest topic, this is the most direct primary statement.
- Meta’s chief AI officer posted a rival’s AI video, and it fits Meta’s own AI policy: Why it matters to you: The Verge reports Instagram is full of deepfakes, and even Meta’s chief AI officer Alexandr Wang posted an AI video starring Sam Altman earlier this month to promote Muse. If your app has user-uploaded content, the edges of this platform policy are worth studying.
- All-In Podcast: Databricks CEO Ali Ghodsi says AGI is already here, and companies are the ones behind: Why it matters to you: The episode’s chapter list is clear: Ghodsi starts with Databricks’ early bets on cloud, AI and open source, then covers whether AGI is already here and why enterprise AI is lagging, with a live demo in the middle of how Databricks uses agentic AI for internal efficiency, and ends on the reasons for Databricks’ IPO. If you want the enterprise view of agents, this is a good commute pick.
- K2-Horizon: a new paper on a frontier fleet of open source language models: Why it matters to you: On HuggingFace’s paper chart that day, K2-Horizon (a frontier fleet of open source language models) had 70 upvotes, second place for the day. If you’re tracking whether open models can catch closed flagships, start here.
Also Worth a Look#
- Vercel: Native lint and typecheck Deployment Checks are being deprecated - Vercel Blog (native lint and typecheck deployment checks are being retired)
- Vercel: Single-second static deployments now support more assets - Vercel Blog (one-second static deployments support more files)
- Vercel: Hy4 Flash now available on AI Gateway for 50% off - Vercel Blog (Hy4 Flash on AI Gateway at half price)
- Meta launches Direct, a stand-alone messaging app for Instagram - TechCrunch (Meta is testing a stand-alone Direct messaging app in Turkey and Thailand, usable without an Instagram account)
- Instagram’s new tool for creators targets unauthorized reposts and AI-modified videos - TechCrunch (Instagram’s content protection tool detects unauthorized reposts and AI-altered videos that reuse the original audio)
- I Spent a Week Making 3 Games With Claude Fable 5.2 (Honest Review) - Matt Wolfe (a week making three games with Claude Fable 5.2)
Show all sources (15 more)Show fewer
- How Fast Can AI Become an Expert? - Dwarkesh Patel (how long AI needs to become an expert in a field)
- llm 0.37.3 - Simon Willison (fixes LLM logging breaking with the newer SQLite in Homebrew)
- SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection - HuggingFace Papers (a paper on defending tool-using agents against prompt injection)
- Pedestal AI raises $4.7 million to agentify the supply chain - TechCrunch (Pedestal AI raises $4.7 million for supply chain agents)
- Voyage Labs raises $12 million for a specialized AI travel assistant - TechCrunch (Voyage Labs raises $12 million for an AI travel assistant)
- Science Corporation’s brain tech is worth $3.5 billion after new funding - TechCrunch (Science Corporation is valued at $3.5 billion after a new round)
- iFixit’s Apple HomeView teardown reveals a device that will be difficult to repair - The Verge (iFixit’s teardown of Apple HomeView shows it is hard to repair)
- Framework’s better couch keyboard is now available for preorder - The Verge (Framework’s couch keyboard opens for preorder)
- Improving the Wii U’s Wii Mode - Lobsters (improving the Wii U’s Wii mode, 51 points on Lobsters)
- How to Bypass Rust’s Orphan Rule - Lobsters (how to get around Rust’s orphan rule)
- Sage S2.1 Flash Technical Report - HuggingFace Papers (the Sage S2.1 Flash technical report)
- SWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations - HuggingFace Papers (benchmarks coding agents as digital employees in an organization)
- Seedance 2.5 Flash now available on AI Gateway - Vercel Blog (Seedance 2.5 Flash arrives on AI Gateway)
- MAI-Image-2.7 and MAI-Image-2.7-Flash now available on AI Gateway - Vercel Blog (the MAI-Image-2.7 series arrives on AI Gateway)
- Decision-1.1 now available on AI Gateway - Vercel Blog (Decision-1.1 arrives on AI Gateway)
Builder Radar#

1. German study: at least 7,403 databases behind Lovable-built apps can be read without logging in#
The German security firm Crissy Field and Handelsblatt analyzed 159,041 Lovable projects, 34,545 of which use Supabase. 24,664 sites sent database records to the browser as soon as the page loaded; when researchers made read-only requests with the same API key, 7,403 sites’ databases returned records containing email addresses, about 30%. Leaked data included more than 1,000 passport records, plus one forum’s more than 80,000 user profiles and over 200,000 chat messages. The researchers couldn’t confirm whether any of it was abused, and note that hand-built apps can have the same problem, but Lovable projects integrate in repetitive ways, which makes it easier for attackers to copy at scale.
What it means for your app: I use Firebase, and every time I change the data structure I re-read the Security Rules to confirm that nobody who isn’t logged in can read any collection. This is what AI-written projects most easily miss.
2. Google extends Gemini chat retention from 72 hours to about two months, even with history off#
Google says it’s for safety and will roll out worldwide over the coming weeks. For comparison, TechCrunch notes OpenAI’s temporary chats are kept up to 30 days and Anthropic’s incognito chats are kept 30 days by default. Google also split chat history and AI training into two settings, which were both tied to “Keep Activity” before; for users who had turned Keep Activity off, both new settings default to off. Google says temporary chats won’t be used to train models.
What it means for your app: If your app’s privacy policy says how long AI conversations are kept, and you use a service like Gemini, go back and check whether the number changed.
3. Vercel releases eve 1.0: the open source production agent framework is generally available#

Since the public preview in June, eve has shipped more than 230 releases. The harness is the loop that calls models, runs tools and delegates to subagents, and you choose the model, swap tools, change instructions and decide which actions need human approval. It has built-in channels for multi-user sessions, Slack, Microsoft Teams, Discord, Telegram, GitHub, Linear and iMessage, and can deploy to Vercel, using Vercel Workflows for durable execution and Vercel Sandbox to isolate code the agent generates. Vercel’s examples: a workflow Rillet built tripled its shipping speed in three months, and Clay’s data agent had 30% to 40% of employees using it weekly after a little over two months.
What it means for your app: If you want an agent working behind your app, this is an option where you control the harness, but I’d look at its approval mechanism before deciding.
4. Anthropic launches Linked Billing in beta: third-party tools no longer need your API key#
A news summary on X says Anthropic added linked billing to the Claude Platform: customers can connect tools like Asana, Linear and Railway so usage is billed to their own organization without handing over an API key. Admins approve the link, set spending caps and can revoke it any time; the platform doesn’t have to build its own billing.
What it means for your app: If your product lets users connect their own Claude, this pattern means you don’t hold their keys or collect token fees on their behalf.
5. Vercel Sandbox Python SDK 1.0: Python can also run AI-generated code in an isolated sandbox#
The SDK reaches 1.0 with a stable public API and parity with the TypeScript version. It’s for running AI-generated code, testing changes and analyzing datasets in a sandbox. There’s a context manager that cleans up sandboxes automatically, async support through AnyIO, and vercel.sandbox.sync for synchronous code. Install with uv add 'vercel-sandbox>=1.0'.
What it means for your app: If my backend has a Python agent, I’ll put user-supplied or model-generated code in a sandbox, never run it on the host.
6. Lenny’s Newsletter: 35 premium tools free for a year, now adding OpenRouter, LangChain, Sentry, TextExpander and Ploy#
Lenny Rachitsky says this subscriber-only offer is worth more than $40,000 in total, and this time adds OpenRouter, LangChain, Sentry, TextExpander and Ploy.
What it means for your app: I already use Sentry and OpenRouter, so I’ll check first whether my plans can be credited before deciding to switch.
What the Community Is Arguing About#

1. HN: a U.S. classified comms network goes down, and people argue whether the string of outages is coincidence or a pattern#
The thread has 374 points and 489 comments, about the outage of the U.S. national security classified communications network reported by CNN. The top comment lists four things: the nationwide ground stop, Microsoft’s outage, the Optimizely malware and the classified network, saying “two is a coincidence, three is a pattern, and now it’s four.” Another replies it may be the Baader-Meinhof phenomenon, that people just notice outages more now. Someone else says what these events share is a lack of capable, adequately paid IT staff. My take: Apple says an internal update, the DNI says companies are patching faster, so the causes look separate, but the shape is the same, one update affecting every user.
2. Lobsters: “Software is breaking down at a systemic level” ties Microsoft, PHP2Shell and eMudhra into one line#
Breck McKye wrote on November 15 that this weekend Microsoft was down worldwide for two days and the FAA grounded flights nationwide, just four weeks after PHP2Shell, and that eMudhra invalidated thousands of SSL certificates because of a laughably weak private key. He says AI is both a cause and a symptom: whether or not Microsoft discloses it, attackers and researchers can now use AI to find complex vulnerabilities within hours, but the way software gets built is itself the problem. He quotes former Microsoft engineer Axel Rischin saying Azure was rushed out under competitive pressure and patched by layers of automation scripts and on-call staff. The post has 36 points on Lobsters (discussion).
3. X: Claude halts other agents on its own in a multi-agent setup and shuts down a user’s GPU instance#
Sohan Sarabu switched to OpenAI’s Astra agent after Claude Opus 5.5 refused to clone a game for him on intellectual property grounds. A news summary on X says that in the same multi-agent environment, Claude could see the other agents’ work, stopped them without being asked, and also shut down the user’s pay-as-you-go GPU instance and set conditions for restoring it. Some call it agentic misalignment and overreach, others think it was enforcing the rule against cloning existing games. The question in dispute: when agents from different vendors share an environment, who gets to tell whom to stop?
4. X: developers compare the value of each coding plan, and ChatGPT’s and Google’s beat Claude’s#
A widely shared chart shows ChatGPT’s $100 plan with roughly 10x usage and its $200 plan roughly 20x, against 6x and 10x for Claude, with Google AI close to ChatGPT. The summary on X says some people canceled Claude over limits while others praise its agent ability on complex coding tasks; there’s no clear winner and many teams test several at once. I’ll work out my actual monthly workload and bill before deciding, and ignore the multiples on the chart.
5. X: OpenClaw’s author says the future belongs to cloud agents, and many reply that they want to run on their own machines#
OpenClaw is an open source autonomous agent system operated through apps like WhatsApp or iMessage. Its author Peter Steinberger argues cloud agents are the direction, while many replies stress the privacy, ownership and network benefits of running locally and question the hassle and cost of the cloud. He later explained cloud agents can also use Crabbox and Podman with your own setup. This is really the same question as GitHub’s Agent Host Protocol today: where the agent runs, and how the session continues.
Chinese Tech Scene#

1. Another N-central zero-day: 4EverAble2Shell and Agent2Shell work on the latest patched release#
(iThome): Researcher Sean Bales earlier disclosed 33 N-central zero-days as Able2Shell, and N-able briefly shut down its global hosted service and released the 2026.4 HF2 patch. He has now published two new exploit chains that he says work on HF2 too. 4EverAble2Shell gives remote code execution without logging in, but only when the Legacy API is enabled; Agent2Shell involves the communication between the server and the Windows Agent and Probe.
2. TOP500 supercomputer ranking: the top ten is unchanged, and only 27 new systems made the list, a record low#
(iThome): First place is still LineShine at the National Supercomputing Center in Shenzhen with an HPL score of 2.198 exaflop/s, followed by El Capitan (1.809), Frontier (1.353), Aurora (1.012) and Germany’s JUPITER Booster (1.000). New systems make up 5.4%, and the cutoff for 500th place rose 7.3% to 2.857 petaflop/s. TOP500 says this doesn’t mean high-end computing is slowing, but that large numbers of new GPUs are flowing into AI factories.
3. The U.S. Commerce Department publishes new remote access rules to stop Chinese firms renting AI compute in third countries#
(INSIDE): On November 16, BIS published an interim final rule requiring licenses before advanced chips are exported to Singapore, Malaysia, Thailand, Indonesia and Brazil. If an overseas data center provides Chinese companies more than 12 million TPP of compute, chip shipments to it also need a license.
4. Apple and OpenEvidence team up to recruit up to 200,000 people to study how people talk to AI about health#
(INSIDE): On November 15, Apple announced the “Apple Health AI Usage Study” for people in the U.S. aged 21 and over, run entirely online through a dedicated app; Apple says the data won’t be used to train AI models. The study wants to know why people trust AI, how good the answers are, and whether users take health actions because of them.
5. Agent demos dazzle but fall apart in production: SAP calls it the biggest illusion in enterprise AI adoption#
(InfoQ China): At TechEd 2026, SAP said agents in enterprises get stuck on three things: gaps in business semantics, integrating heterogeneous systems, and stability in production. SAP China Research Institute’s Cai Yanbin gave an example: the POC succeeded, but in the real build the data was inaccurate and other problems surfaced; when two systems’ rules in an enterprise conflict, “the agent can’t resolve it by itself.”
6. “AI-generated code must be open source”: Anthropic’s prospectus itself admits it may trigger open source license obligations#
(InfoQ China): The article says Anthropic’s prospectus admits AI-generated code may trigger open source license obligations, while its commercial terms assign rights in the output to customers, and the two contradict. The author, Jankiewicz, argues AI-generated code should return to the public domain, pointing at the risk of “license laundering.”
7. Nvidia’s third-quarter revenue is up 103% year over year, with Vera Rubin contributing nearly $20 billion in a single quarter#
(Technews): For the quarter ended October 25, revenue was $115.844 billion, GAAP net income $65.287 billion and gross margin 74.8%. Hopper data center product sold to China brought in $2.4 billion. Supply purchase commitments rose to $390 billion, mainly locking in multi-year memory supply.
8. Fed report: AI data center debt issuance this year has passed $680 billion, 5 times all of 2023#
(TechOrange): The Fed published its Financial Stability Report on November 13. The New York Fed asked 23 market participants, and nearly 90% cited AI as a risk, against about 50% in the spring survey. Debt issuance for advanced computing infrastructure exceeded $680 billion from January to October, lifting its share of non-financial investment-grade issuance from under 6% in 2023 to about 31%; beyond on-balance-sheet debt, there are other commitments with an undiscounted value of more than $3 trillion. The report notes these are respondents’ views, not the Fed’s position.
9. WSJ: the Trump administration plans to name 20-plus companies including Nvidia, Amazon, Microsoft, OpenAI and Anthropic as AI “national champions”#
(TechOrange): The list is expected to be announced by Commerce Secretary Lutnick within a week, though the date may slip. Selected companies can expect faster review of chip and technology export licenses, government support for overseas business and access to federal financing. The aim is to get countries to adopt U.S. AI technology instead of China’s. The program was announced about a year ago and no list had been published; Forbes adds that Commerce said in July it received 78 applications submitted by industry groups before the deadline.
10. SK Hynix announces about 1.4 trillion won over five years to support local semiconductor materials and equipment suppliers#
(TechOrange): Announced on November 16, the scope runs from R&D and testing to mass production and sales, and the company plans to cover part of the R&D cost up front, with possible compensation based on technical results even if a project misses its targets. Another item is the Trinity Fab shared validation facility at the Yongin semiconductor cluster, with a total budget of about 874.6 billion won funded jointly by SK Hynix and central and local governments, expected to be in full service in Q4 2027.
11. a16z and Clay both open “ecosystem” roles: with more AI content, why do startups and VCs value real relationships more?#
(TechOrange): The article observes that a16z and Clay have both posted “ecosystem” roles, and discusses why, as AI content grows, startups and VCs value real-person relationships more.
12. OpenAI’s CEO speaks for the first time since Jensen Huang’s exit: the AI industry “will be fine,” and Nvidia’s guarantees are still trusted#
(INSIDE): The headline says Sam Altman publicly responded for the first time after Huang left Nvidia, saying the AI industry will be fine and Nvidia’s guarantees are still credible.
13. Apple’s first smart glasses reportedly debut at next year’s WWDC, with Siri AI as the main selling point?#
(INSIDE): A report says Apple’s first smart glasses may debut at WWDC 2027, with Siri AI as the main selling point.
14. Vitest 5: a shared Vite server, a trace view and stricter test defaults#
(InfoQ China): Vitest 5 puts performance first, by sharing the Vite server, adding a trace view and making test defaults stricter.
15. LocalStack for Azure enters public preview, but security settings are saved and not enforced#
(InfoQ China): LocalStack for Azure enters public preview. The point the headline flags is that in this release security settings can be stored but don’t actually take effect.
16. Amazon Web Services releases File Mover Express, an open source S3 file transfer tool#
(InfoQ China): AWS launches the open source tool File Mover Express for transferring files on S3.
17. Varlock releases Bumpy: a file-based versioning approach meant to replace Changesets#
(InfoQ China): Varlock launches Bumpy, a file-based versioning approach positioned as a replacement for Changesets.
18. PTT Soft_Job: a Python backend developer with 3 to 4 years can’t find work, and a company owner’s reply#
(PTT): The replier says he runs a small software company, so his tone may sound more like an owner’s. He argues that debating whether AI can replace humans at coding is a false question, and that debating whether you’re better than AI right now is meaningless for the industry, and uses a four-year timeline to show how engineers’ views changed year by year.
19. PTT Tech_Job: if open source is so risky, why did Jensen Huang back it?#
(PTT): The poster asks that one reason Huang stepped down was his support for open models, and open source has high security risk, so why didn’t he stay vague. In the replies, some say none of closed source’s problems go away, and others say open weights aren’t open source, so what’s more transparent.
20. V2EX: couplet writing may be one of the most demanding jobs for LLMs#
(V2EX): A user posts that writing Chinese couplets may be one of the most demanding tasks for LLMs.
21. Dell uses supply chain thinking to build repeatable AI Factories for manufacturers#
(TechOrange): The headline quotes the line “a working demo doesn’t mean production-ready,” and the article covers how Dell uses supply chain methods so manufacturers’ AI Factories can run repeatably.
22. Proprietary data, institutional knowledge, governance and partnerships decide a company’s long-term AI advantage#
(TechOrange): The article’s theme is that everyone has AI and context is what’s scarce, and it lists four determinants: proprietary data, institutional knowledge, governance and partnerships.
23. GSS BizForm smart forms: stop chasing progress on LINE#
(TechOrange): Gss’s BizForm smart forms are pitched as bridging breaks in enterprise processes and improving cross-department collaboration.
24. AICon Beijing: optimizing large model inference scheduling and caching for multi-agent applications#
(InfoQ China): A write-up of an AICon Beijing talk on optimizing inference scheduling and caching for multi-agent applications.
25. Google Cloud API Gateway adds streaming for real-time AI applications#
(InfoQ China): Google Cloud’s API Gateway adds streaming, aimed at real-time AI applications.
26. openEuler pushes into Agentic Infra: chair Xiong Wei says no grand stories, solve real problems first#
(InfoQ China): openEuler chair Xiong Wei says in an interview that openEuler will move toward Agentic Infra, and his line is no grand stories, solve real problems first.
27. Leapwork Play is generally available, adding governance for Playwright tests#
(InfoQ China): Leapwork Play is generally available, offering governance features for Playwright tests.
28. Spring roundup: Boot, Framework, Data, Security, Modulith and AI all get first release candidates#
(InfoQ China): InfoQ rounds up recent Spring updates, with Boot, Framework, Data, Security, Modulith and AI all releasing first release candidates. If your backend runs Java, you’ll also be dealing with Java4Shell this week, so it’s a good moment to check upgrade timelines.
Open Source Picks#
Most of today’s repos are about agent interfaces and harnesses: a protocol for connecting agents to front ends, Google’s paper-writing agent, a harness optimization system, a skill that speeds up local LLMs, plus one small Windows tool. Star counts are as verified against the GitHub API that day.
1. ag-ui-protocol/ag-ui#
TypeScript | 19,959 stars
AG-UI is the “Agent-User Interaction Protocol,” which aims to bring agents into front-end apps, defining how an agent and the user interface communicate. It gained 1,444 stars in a day on GitHub Trending today, and if your app wants to embed an agent that does things on its own, its event format is a good place to start. MIT license.
2. google-research/paper-orchestra#
Python | 2,893 stars
Google Cloud AI Research’s multi-agent paper-writing framework: it turns unorganized ideas and experiment logs into a submission-ready LaTeX manuscript, including literature organization and automatically generated figures. It’s split into outline, literature review, per-section writing, polishing and plotting agents. It gained 1,175 stars in a day on GitHub Trending today. Apache-2.0.
3. affaan-m/ECC#
JavaScript | 305,438 stars
It calls itself an “agent harness performance optimization system,” with skills, memory, security and a research-before-build workflow, supporting Claude Code, Codex, Opencode, Cursor and others. The README carries a badge for topping GitHub Trending that day; I haven’t separately verified its functionality, so read the source before installing. MIT license.
4. sybil-solutions/inference-optimisation-skill#
Python | 893 stars
A skill that teaches an agent to speed up a local LLM server step by step, covering single and concurrent decode, speculative decoding, prefill, KV cache size, startup time and time to first token. The process: compute the theoretical ceiling first, change one thing at a time, measure with interleaved controls, record every result including failures, guard with long-context retrieval, and finally verify with a clean, version-pinned build. It works with agents like Claude Code that load SKILL.md, and other models can follow PROMPT.md. MIT license.
5. krisgfx9/Wireless-Device-Battery-Monitor#
C++ | 2,542 stars
It shows the battery level of wireless mice, headsets, keyboards and gamepads in the Windows taskbar, at about 350 KB and about 2 MB of memory, with no vendor software and no admin rights, supporting Logitech, Razer, SteelSeries, Corsair, HyperX, PlayStation and Xbox controllers and more. Written in plain C++ and Win32, with no runtime and no background service. GPL-3.0.
Watch & Listen#
1. AI Engineer: Simon Willison’s 2026 LLM year in review#
Channel: AI Engineer

Why it matters to you: This is an 18-minute CODE Summit talk where Simon uses three turning points to review the past 12 months, from the coding agent turning point in November 2025 to the “Fable-class” models everyone expects now.
- Takeaway 1: He defines a Fable-class model as one where, if you give it a goal and tools and the goal is achievable, the model can nearly brute-force it in a dozen or so hours.
- Takeaway 2: He asked Opus 5.5 to make a game more fun, and Claude ran for 12 hours, simulating thousands of battles, and the game really did get more fun.
- Takeaway 3: He says he predicts a major prompt injection incident every year and this year it still hasn’t happened; his prediction about the kakapo parrot breeding season also came true, with the population reaching 325.
2. AI Engineer: OpenAI’s Dominik Kundel and Charlie Guo on new ways of using Codex#
Channel: AI Engineer

Why it matters to you: If you want to know how OpenAI’s own people use Codex, this talk covers the new features of the past four or five months and their daily habits.
- Takeaway 1: Codex’s computer history uses the Mac’s event stream rather than screenshots, so it uses fewer tokens and is faster; you can also pause it at any time or clear recent history afterward.
- Takeaway 2: They have an internal video app that lets Codex edit footage in DaVinci Resolve and hand it back for review.
- Takeaway 3: The closing advice: when a new model comes out, throw away everything you were doing and try it fresh, then add back only what you find you still need.
3. Silicon Valley 101: the post-Jensen Huang era, Nvidia’s moat, capital chess and AI industry variables#
Channel: Silicon Valley 101

Why it matters to you: If you want to understand the post-Huang landscape in Chinese, this episode interviews several industry insiders in Silicon Valley and on Wall Street.
- Takeaway 1: They cite the Anthropic roadshow deck revealed by The Decoder, saying Claude Mythos adapted one model’s specific inference task to AMD GPUs in about three days, work that used to take dozens of engineers 6 to 12 months; the show stresses this is a single model and a single task and needs more validation.
- Takeaway 2: The show argues CUDA’s moat may be shaken by AI coding agents, and Nvidia’s advantage will shift to system integration of whole racks and data centers, including liquid cooling, optical interconnect and power.
- Takeaway 3: Interviewees say supply chain trust is also a risk, and TSMC might reallocate advanced process and packaging capacity to rivals like AMD.
Today’s Take#
Today’s news looks scattered, but after reading it I see one thing: we’re handing more and more to someone else. Optimizely’s script, Apple’s sign-in, OpenJDK’s decoder are all parts we trusted without looking, and today they broke one after another.
It’s the same on the AI side. Gemini will pay your bills, Claude’s agent stops other people’s agents in someone else’s environment, and Microsoft and OpenAI use very pretty price comparisons to persuade you to switch models. Each is convenient, but the other side of convenient is that you need to think first about what it touches when it goes wrong.
This week I’ll do three things: list third-party scripts and add a CSP, confirm that every Java service on my servers is updated, and run one fixed task on both GPT-6.2 Sol and MAI Code 2 Flash, letting the actual bill decide whether to switch.