Protip: Every time you read a headline of the form "X plans to invest $Y in Z", these generally have zero contractual force (e.g. "Nvidia plans to invest $100B in OpenAI"). The "plan" is usually implemented as a nonbinding letter of intent, where there would be zero recourse if someone backed out.

It's immensely charming that Wikipedia still contains deeply technical articles that have, like, some rando's handwritten notes as the reference image for how an algorithm should work.

It's a little bit silly that if you're a billing administrator in Azure, you can list all the resources that went into a particular subscription, but you can't `az resource list`. I get the distinction and understand why it might be relevant, but nevertheless it seems like quite a pain.

PSA: If you received a notice from GitHub about participating in the "GitHub Developer Fund", it's a scam. The repository has been blocked by GitHub at this point, but because the scammers directly mentioned visible/successful developers, they'll be more likely to get into your inbox.

Can you imagine trying to build the Hoover Dam without Linear or the Apollo missions without git?

"FTC Probes Whether Ticketmaster Does Enough to Stop Resale Bots" Narrator: They don't.

Signs your laptop may be too big: the cases that look like they'll fit best are marketed as "portable monitor" cases rather than laptop cases.

Trying to understand the median American person through the lens of DC politicking is like trying to understand the median Chinese person through the lens of Beijing's activities. It's just a really terrible proxy for each nation as a state, as a culture, and as a people.

I don't care if it's too late to stop spotted lanternflies as an invasive species overall. Every delay of their spread gives natural predators more time to adapt and take them on. Nice work @stevenljohnson.org!

Cake by the ocean?

✈️ I'll be in Dallas this week to give a talk on lessons learned from building AI platforms at scale. If you're around the DFW area this week and building things with software in any capacity, I'd love to say hi and swap notes over a morning coffee somewhere. DM me!

You know how occasionally a third-party browser extension owner will get bought out by some data-harvesting business? Cannot *wait* to see the version of that for MCP extensions: chromewebstore.google.com/detail/brows...

Helpful LLM: I wrote a prompt, spawned two subagents, and 204 seconds later Cline has trawled BLS datasets, extracted the correct dataset, rendered a chloropleth, and uploaded it to S3 for me.

I refuse to stop using em dashes just because everyone thinks that's how AI writes. It's a perfectly good piece of punctuation—and I won't be deterred. "Why should I change? He's the one who sucks." -- Michael Bolton, _Office Space_

The Pandemic Mitigation Collaborative currently estimates that about 1 in every 55 people (~1.8%) is infectious with SARS-CoV-2. Perfect timing for everyone going back to school.

Current status writing my own kernel drivers:

Windows 95 is three decades old this month. From left to right, here are the floppy stacks you could have used back then to install MS-DOS, Windows 95, and Microsoft Office respectively. For the current generation of Windows, it would take approximately 3,600 floppies to install.

A riddle for you: What do "cyber hornets", "digital energy", and "apex property" all have in common? Answer: They're all phrases one public company CEO has used to describe Bitcoin. In this thread I'll explore a persistent financial puzzle: how markets value unusual assets. But first: a story.

Anthropic presents an 11% attack success rate as a victory: "When we added safety mitigations to autonomous mode, we reduced the attack success rate of 23.6% to 11.2%." Isn't "one out of every nine prompt injection attacks succeed" objectively horrifying, and not something to celebrate?

NYT's investors, Fivespan Partners, are pressing for more AI. While it's very common to imagine expanding subscribers with ML techniques (e.g. identifying clusters of people who don't renew), it'll be a nightmare if this is used as a pretext to find ways to push generative AI slop on readers.

Elon Musk (through xAI) filed a 61-page lawsuit against Apple and OpenAI today. As a technologist with a lot of clients relying on clarity about what's happening in the fast-moving generative AI space, I try to pay close attention to this stuff. Here's a thread of some of the interesting tidbits:

What is the line between #dadjoke and merely being a pun?

In vendor conversations, you should be deeply skeptical of any marketing claim of a truly novel result, capability, or outcome that's driven by LLMs. Generally speaking, vendors do not have access to step-function-better models compared to what's commercially available.

One of the best and most informative pieces on deep AWS knowledge from (who else?) @quinnypig.com: www.lastweekinaws.com/blog/aws-in-...

A new and deeply unasked-for behavior in LLM responses (particularly OpenAI's ChatGPT) is the "Let me know if you also want to..." suffix that gets tacked onto everything. It's a shameless prod to burn more tokens and I find it greatly annoying.

Spot the difference in Eric Schmidt (February 28, 2025) vs. Eric Schmidt (August 19, 2025): (h/t: @himself.bsky.social)

Ea-Nāṣir, you son of a gun.

Likewise, if you don't test for COVID, the COVID numbers look great!

Whenever you're having a rough day at work, just remember that an OpenAI executive somewhere greenlit the use of this horrific abomination of a bar chart:

One of the lofty software engineering ideas from the 1980s was "literate programming" (from Knuth in 1984 [0]). It failed because it's hard to write code that way; it's a lot of effort. But ... could generative AI play a role in giving it a second life? www.literateprogramming.com/knuthweb.pdf

A great deal of these "I replaced X employees with AI" thinkpieces give off the same energy to me as "I lost 20 pounds by amputating a limb".

Some days I think I'm raising intelligent and empathetic human beings. Other days, like today, they're standing around an anthill chanting "EAT! THAT! ANT! EAT! THAT! ANT!" at each other. C'est la vie.

Congrats to the International Physics Olympiad winners, the US team. These bright minds each won a gold medal (shown left to right in picture): Feodor Yevtushenko, Brian Zhang, Joshua Wang, Allen Li, and Agastya Goel.

One of the dumbest things in English is that we have decided to use the word "cousin" to mean completely different relationships. Notice that here your "first cousin once removed" can be _either_ your cousin's children or your grandparent's sibling's children.

Delivery bots in LA. #resist #clankers

It's deeply frustrating to see our nation's research and monitoring capacity dim day by day. Will the last reputable scientist leaving please turn out the lights? michaelrlowry.substack.com/p/navy-set-t...

My children have exactly three weekend operating modalities right now: * "Let's build Legos!" * "Let's go to the pool!" * Watch KPOP DEMON HUNTERS for the 73rd time, and if I don't sing the Saja Boys part at full volume they get mad at me.

This is chlorine trifluoride, ClF3. It's extremely reactive, corrosive, and hypergolic. "Hypergolic" means it has such a strong affinity for electrons that it vacuum-oxidizes things you wouldn't normally consider flammable, like sand, asbestos, or humans who don't follow safe handling procedures.

One of the more fun reads circulating these days is "AI 2027" (ai-2027.com). In it, the authors describe a scenario that could easily be the plot to a thriller novel: LLM advances lead to runaway AGI and humans are no longer driving humanity. Their timeline for this is ~2030.

Asking the OpenAI Whisper audio model to transcribe 30 seconds of silence generates the output "ترجمة نانسي قنقر", which is "translation by Nancy Qanqar". Nancy Qanqar is a volunteer translator of (among other things) many pirated movies, so that invites an interesting conclusion on data sources.

ln(😅) = 💧ln(😄)

Grok 4 appears to take into account Elon Musk's personal opinion on things when asked questions, as in this example via @simonwillison.net:

Some falsehoods CIOs seem to believe about AI that mystify me: * "Has AI" means "uses an LLM". * You don't need to know anything about how LLMs work in order to use them effectively. * Anything a computer can do, an LLM or agent can do just as well, with the right prompting and tools.

This deal makes a lot of sense given that a large part of the American cereal aisle basically *is* candy.

What if every time you went to thispersondoesnotexist.com, it wasn't actually using generative AI but rather causing a real person's instantaneous and simultaneous erasure from this timeline?

My most problematic take is that if a coffee drink comes from a machine with a touchscreen, it's 1,000% undrinkable trash.

This is Sungjin Cho, the person that the US DOJ says is a cybercriminal who co-masterminded a profitable hack of the SEC EDGAR system:

It's fascinating how much behavior in humans and animals is explained by "evolution optimizes for transmitting genes, not for health or civilized society".

[Client → Server] PSH, ACK │ Payload (17 bytes): 0xF0 0x9F 0x91 0x8B 0x20 0x48 0x69 0x2C 0x20 0x42 0x6C 0x75 0x65 0x73 0x6B 0x79 0x2E 👋 Hi, Bluesky.