Phil Steitz

Let me ask the question that is going to have me burn on the pyre: why do we need models to learn good representations already? Why does it matter?

My take on possible oil scenarios & conflict in Middle East. Inspired by Jurassic Park and Jeff Goldblum’s explanation of chaos theory. This one hit close to home as I know too well tail risks. My mom cried - should have warned her first. Lots of humility amidst uncertainty.

I really hope that this does not turn out to be a Claude targeting mistake. We will likely never know, but this is precisely the reason that Dario did not want Anthropic tech used without humans in the loop.

Wow

generative-ai/gemini/agents/always-on-memory-agent at main · GoogleCloudPlatform/generative-ai Sample code and notebooks for Generative AI on Google Cloud, with Gemini on Vertex AI - GoogleCloudPlatform/generative-ai

This looks interesting. I have always hated vector databases because they force stock embeddings.

github.com/GoogleCloudP...

A little faded, but still there.

I thought about taking the flag down today out of shame. But then I worried about what my neighbors would think - not in the way one normally uses that phrase. I was afraid that they would think that I had given up. I am not going to give up on my neighbors.

The Illusion of Building AI makes it dramatically cheaper to produce software that appears to work. But 'building an app' and 'engineering a system' are two different activities that people keep confusing, and the gap between...

uphack.io/blog/post/th...

This is excellent.

A Comparison of DeepSeek and other LLMs Recently, DeepSeek has been the focus of attention in and beyond the AI community. An interesting problem is how DeepSeek compares to other large language models (LLMs). There are many tasks an LLM...

www.tandfonline.com/doi/full/10....

Saw this in dead trees version which is ancient by now, but the approach is cute. Claude Sonnet 3.5 smokes same epoch other models on 0-shot slop classification. Error rates are not impressive though.

Welcome | Apache Otava

otava.apache.org

Great to see leading performance engineering tools coming to the ASF

Rattler security

After clearing 2 rodent nests there, we were actually happy to find “security” under the AC unit yesterday. Loud rattle is a plus.

When I get bored with science, it's because I've read the same paper three times from different authors whose media diet consists of corporate PR releases and unhinged slop summaries for viral arxiv drops.

I don’t think it is so much devaluing as over-simplifying. Claude can do amazing things with code because the language is so simple and good prompts can be translated into that low-dimensional space easily. The mistake is thinking that, like (some) code geniuses, that must mean it can think.

It’s hard to image any leadership job in any company where Trump’s no plan, grab-ass leadership style would not get you fired. This is so embarrassing.

Go work with Ted. You will learn something.

AI could prevent construction delays before they happen What if a construction project could rewrite its own schedule the moment a problem appears? A new peer-reviewed study from the University of East London (UEL) suggests that artificial intelligence could make this possible—detecting emerging risks and automatically adjusting project plans before delays spread across a site.

Integrating AI-driven risk detection with scheduling systems could enable construction projects to automatically adjust plans in response to emerging issues, potentially reducing delays and improving resilience.

I am tempted to try the problem :)

Beautiful day for a run

The Man Who Stole Infinity | Quanta Magazine In an 1874 paper, Georg Cantor proved that there are different sizes of infinity and changed math forever. A trove of newly unearthed letters shows that it was also an act of plagiarism.

www.quantamagazine.org/the-man-who-...

This article really bugs me. Missed opportunity to explain the arguments. Smear piece distracting from what must have been a great collaboration. I expect more from Quanta.

Cognitive Debt: When Velocity Exceeds Comprehension | rockoder A systems analysis of how AI-assisted development creates a gap between output speed and understanding, and why organizations cannot see it happening.

www.rockoder.com/beyondthecod...

Verified Spec-Driven Development Verified Spec-Driven Development. GitHub Gist: instantly share code, notes, and snippets.

gist.github.com/dollspace-ga...

We can actually do this kind of thing now. Let’s harvest the AI surplus to improve software quality.

How to delete your account | OpenAI Help Center Delete your account by submitting a request to our Privacy Portal or within ChatGPT directly.

help.openai.com/en/articles/...

Any sufficiently large k-nn is indistinguishable from magic
🧙‍♂️

When a company in an industry built on hype tells you that a use case is a bad idea—and actually dangerous—that means it’s a *catastrophically* bad idea.

While whiskey Pete plays army with Grokk and shoots at speedboats, our defenses against today’s actual imminent threats fall into disrepair.

The U.S. spent $30 billion to ditch textbooks for laptops and tablets: The result is the first generation less cognitively capable than their parents | Fortune Neuroscientist Jared Cooney Horvath said older generations “screwed up” giving students access to so much technology: “I genuinely hope Gen Z quickly figures that out and gets mad.”

Now THAT's a headline.

"The U.S. spent $30 billion to ditch textbooks for laptops and tablets: The result is the first generation less cognitively capable than their parents"

fortune.com/2026/02/21/l...

Did ChatGPT write this talking point for you, Sam?

Or do you just *organically* suck this bad?

By the numbers Here’s where the cryptocurrency industry stands, early into the 2026 election cycle: $288 million: Total cryptocurrency industry spending toward the 2026 election cycle to date. This includes funds sent to pro-crypto PACs, direct contributions to candidates, and contributions to non-crypto PACs.a $221 million: Cash on hand with pro-crypto super PACs, ready to deploy in the midterms $100 million: Additional committed funds that pro-crypto PACs say they have secured but haven’t yet appeared in FEC filings $3 million: Already spent by pro-crypto super PACs in the 2026 cycle, primarily on special elections in Florida and Virginia $74 million: Contributions to Trump PACs in the 2026 cycle by crypto companies and executives

Here’s where the cryptocurrency industry stands, early into the 2026 election cycle:

• $288 million: Total cryptocurrency industry spending
• $221 million: Cash on hand with pro-crypto super PACs
• $100 million: Additional committed funds
• $3 million: Already spent by pro-crypto super PACs

Symmetry in language statistics shapes the geometry of model representations Although learned representations underlie neural networks' success, their fundamental properties remain poorly understood. A striking example is the emergence of simple geometric structures in LLM rep...

In our new preprint, we explain how some salient features of representational geometry in language modeling originate from a single principle - translation symmetry in the statistics of data.

arxiv.org/abs/2602.150...

With Dhruva Karkada, Daniel Korchinski, Andres Nava, & Matthieu Wyart.

The Promptware Kill Chain - Schneier on Security Attacks against modern generative artificial intelligence (AI) large language models (LLMs) pose a real threat. Yet discussions around these attacks and their potential defenses are dangerously myopic...

www.schneier.com/blog/archive...

Behold: the first-ever list of news outlets that have banned generative AI in their reporting. As of today, this is literally information that you cannot find on Google.

My goal is to fill the starter pack, so please send over suggestions with supporting evidence!

go.bsky.app/8cn1XfT