Kevin Webb

(I was trying to figure out the compressive strength of Type S mortar as it cures and Claude started trowing up questions about the structural integrity of the building just to take me down a peg...)

Provide critical feedback on prompts, and suggest ways to improve my understanding of topics. Do not provide positive feedback or praise for my questions. Focus on responses that help me understand the problem and how to further investigate answers.

the foundation is solid

idk, I think this is mostly a design choice on the part of the bot purveyors. Sycophancy sells.

Putting simple instructions into the context turns Claude (and ChatGPT to a lesser degree) into a combative critic that strawmans wild failure scenarios just to poke holes in my questions...

Why aren’t these everywhere?! Was in a German supermarket last year and found *giant* bags of them for almost nothing. I seriously considered checking another suitcase just to bring them back…

Airlines have been trying to use computers to decommodify and differentially price seats since the time of their invention. But “agentic purchasing” and embedding payment into the chat is going to make past conflicts like “screen bias” look quaint…

jolt.law.harvard.edu/assets/artic...

Why Does Kevin Williams Keep Getting Static For Saying That Chinese Cars Are Good? Western car makers keep thanking him for saying the things their bosses won't believe. Ep. 28

Glad to chat with Alex Roy and Joel Johnson on Tool Or Die last week;

China's car game is solid -- I can understand why some are blindsided and taken aback, but others....I don't know how they didn't see it coming. Give it a listen.

www.toolordie.com/p/why-does-k...

Sinophobic petrostate fusionism

just picked this up yesterday!

Your investigation plan has a scope problem - you're trying to tackle two fundamentally different levels of analysis simultaneously. Sampling methods operate on the final output logits, while cross-layer analysis requires examining internal model states. This混合 will likely lead to confusion about what you're actually measuring.

Hadn't seen this happen before: Claude dropped into Mandarin to explain that I was attempting to mix (混合) two different ideas in my request.

The systems that power GNSS are fascinating: yet another example of DIY infrastructure from a bygone era of global scientific collaboration. They're fragile not only due to geopolitics, but they're designed primarily for academic research workloads, not the modern 24/7 systems using this data.

GitHub - geoceiver/ground-control: A high-performance GNSS data processing and archiving system A high-performance GNSS data processing and archiving system - geoceiver/ground-control

Just posted a first draft of something I've been working on. Geoceiver began as part of an IoT project looking at ways to improve GNSS precision in urban environments. Turned out the data infrastructure powering GNSS is mostly text files shared on FTP sites, and it's holding back lots of R&D.

I'll see your ColdFusion template engine and raise you a couple million lines of Smalltalk:

Per @steveklabnik.com's comment in the podcast, there's no quicker way to realizing LLMs limitations than actually using them. It's amazing to get a boost on stuff like the API upgrade, but you have to keep a firm grip on the wheel, and know where you're going, or things get ridiculous real fast.

One of the best examples of LLM developer tooling I've heard is from a team that supports software from the 80s-90s. Their only source of documentation is *video interviews* with retired employees. So they feed them into transcription software and get summarized searchable notes out the other end.

How Morgan Stanley Tackled One of Coding’s Toughest Problems The finance giant built its own AI tool to help modernize its legacy code—something it said existing tools on the market still struggle with.

For a lot of orgs accelerating code maintenance is even more significant, as the WSJ notes in this piece today. These are systems with astronomical switching costs that are often holding back all kinds strategic/org-level progress.

Being able to type, "This code uses last year's REST API spec. Can you update it to use the new version defined in this PDF?" I did that last week to exfiltrate data from a not so great SaaS product and had simple UI up and running in minutes rather than days of work.

Using LLMs to help enterprise devs level up from tedious, time consuming plumbing so they can focus on how software/data infrastructure actually adds value is huge productivity lift, not just for developers but the orgs as a whole.

The discussion about enterprise IT and containing cost/fixing underperforming SaaS infrastructure is fascinating.

My biggest home runs with LLMs involve feeding in large ~poorly defined REST API specs and getting boilerplate code/interfaces out the other side so I can build useful stuff on top.

This is such an important conversation and mirrors where I'm at on LLM use/value.

The critique got frozen in amber a 1-2 years ago, and is fueled by (valid) structural issues like energy, copyright, deepfakes, etc. But the tech continues to evolve in ways that are unrelated to these problems.

Great piece on BYD's $8,000 EV.

Also, fascinating tidbit about Shanghai banning small/inexpensive cars:

We’ve Lost the Plot: Sinophobia and the Collateral Damage of American Primacy What Visa Bans Teach America About Itself

"We have entered a feedback loop in which Sinophobia is both a byproduct and a driver of strategic rivalry. It fuels the antagonism, and the antagonism legitimizes the prejudice. And in the meantime, real people — scientists, students, ordinary Asian Americans — are swept up in it..."

Flock Safety

Doubly unfortunate that conversations about surveillance devolve into unrelated fights about red light and speed cameras (which are useful/good).

Companies like Flock, that power nationwide LE infra, have a different definition of "safety" and don't even make products that enforce traffic laws.

That falls apart once you start digging into how the enforcement of laws works in the US. Personally, I think the folks that work(ed) on privacy policy in the US are mostly interested in talking to themselves about interesting legal theories. But issues with 4th/14th are about rights, not privacy.

Unfortunately not enough folks who work with this technology understand how we got to where we are with 4a law's limits or intentions. Can't recommend enough Sarah Seo's book unearthing this history. thewaroncars.org/2019/10/31/t...

Big lesson from the last several years: 4a law isn't equipped to handle the harms embedded in pervasive surveillance tech like ALPR.

Also, lots of people (who don't support LE dragnets) ended up on the wrong side of the right to privacy in public fight because pervasive surveillance seemed useful.

The last remaining point of bipartisan consensus: blaming the rest of the world for a half-century of mistakes.

Er 70s-80s trends like MTBs. US mfrs were already toast by the 90s, which is why they asked for protection.

I don't get bringing up past requests for protection ("We love tariffs🇺🇸🇺🇸🇺🇸"?), the fictional history of decline, or the point of asking for sectoral relief from an admin that could care less.

Bonkers.

Astonishingly, the US bike lobby is attempting to get sectoral tariff relief by blaming a lack of protectionism and "cheap" foreign bike imports in the 1990s for the collapse of US manufacturing.

1) Bonkers logic.

2) US cos failure to respond to 90s trends like MTBs killed the domestic industry.

SoftBank Stargate Venture With OpenAI Hits Snags on Tariff Fears SoftBank Group Corp.’s plans to invest $100 billion in artificial intelligence infrastructure in the US have slowed, with economic risks stemming from Washington’s tariffs holding up financing talks.

OpenAI's problems are way bigger than tariffs, but still...

"But uncertainty is weighing on talks. With tariff discussions on the table for everything from server racks to cooling systems and chips, data center build costs may rise by 5% to 15%, with some operators facing even higher prices."

As funny as this is to watch it looks like it's at least partly caused by the poor over the hood visibility of modern mega trucks. Unfortunately it's not just rocks these trucks run over, and in most crashes, the truck wins.