Today Claude made up a new word: "re-pend" for "to make pending again". I almost just took it at face value.
Alan Buxton
@alanbuxton.bsky.social
Serial CTO etc in tech startups. Based in London. Also sometime classical guitar player. Working on an open(ish) data side project https://syracuse.1145.am
TIL that postgres pg_restore for a single table doesn't also restore its indexes. Nice.
In praise of smaller large language models Part of my side-project's codebase uses Flan T5 XL for some generative tasks. Flan T5 is an LLM that was released in October 2022. The XL version that I use has 3 billion parameters. Might sound like a lot but it's small by modern standards. Kimi K3 has…
In praise of smaller large language models
Part of my side-project's codebase uses Flan T5 XL for some generative tasks. Flan T5 is an LLM that was released in October 2022. The XL version that I use has 3 billion parameters. Might sound like a lot but it's small by modern standards. Kimi K3 has 2.8 trillion parameters. It has been hogging a lot of resources on my machine.
alanbuxton.wordpress.com
"Very informative — and it overturns my earlier assumption." Is what Claude wrote this morning when I told it to run a particular failing regression test. Shows how important tests are. Not even high-faluting evals but good old fashioned integration tests.
Benchmarking Coding Agents on Databricks’ Multi-Million Line Codebase They ran a comprehensive evaluation on our tasks, code base, infra. It's been produced by more than 3,000 software engineers, spans 3 hyperscalar clouds and many languages and tasks. www.databricks.com/blog/benchma...
Check in on Syracuse Progress Syracuse Company News has been working really well on company news (better than the alternatives), but on industry news it hasn’t been so strong. It has been improving so I asked Claude Opus 4.8 to review Syracuse’s performance on industry news since March based on…
Check in on Syracuse Progress
Syracuse Company News has been working really well on company news (better than the alternatives), but on industry news it hasn’t been so strong. It has been improving so I asked Claude Opus 4.8 to review Syracuse’s performance on industry news since March based on the stats in the benchmark that compares Syracuse with Exa, Linkup, Perplexity and Tavily. Here is what it had to say:
alanbuxton.wordpress.com
Today Fable told me: "don't let the LLM derive facts that code can compute." Seems like a good guideline for working in this wacky world of AI.
Fable: “don’t let the LLM derive facts that code can compute.”
Seems like a good guideline for working in this wacky world of AI. The context: This is about my side project, Syracuse Company News ( It's a structured company news provider that you can plug your agent into and be confident you're getting decent quality results with less risk of false positives / irrelevant stories fouling things up compared to other search tools.
alanbuxton.wordpress.com
“We can’t ask AI, it lies” vs “Here is my superpower prompt” It’s well-documented that Gen Z / Gen Alpha are not exactly fans of AI1. And that older professionals can’t live without it. My own 9 year-old kid has told me that games made with AI are terrible and that “we can’t ask AI, it lies” (in…
“We can’t ask AI, it lies” vs “Here is my superpower prompt”
It’s well-documented that Gen Z / Gen Alpha are not exactly fans of AI1. And that older professionals can’t live without it. My own 9 year-old kid has told me that games made with AI are terrible and that “we can’t ask AI, it lies” (in the case we were talking about, it did lie). Marc Andreessen believes he has a superpower prompt that helps him understand anything.
alanbuxton.wordpress.com
My 9 year old boy wants to write a Roblox game. The tutorial gives you an option to do it yourself or to use AI assistant. He didn't want to use AI. Because he's said that he has seen people use AI in games and their games are just, "bad".
My Claude Code editor still tells me that I have Fable included in my plan limits till June 22. But obvs Fable doesn't exist right now. I wonder what happens if Fable gets reactivated on June 23?
Noticed my first use of ad-supported ChatGPT today. They did it really well. I asked a question, got an answer as expected and then, after the answer, there was a small ad related to the topic. Nicely done.
Two Things Can Be True: Anthropic’s new Fable I received two newsletter subscriptions within an hour of each other today regarding Anthropic's new Mythos-class AI model, named Fable. First, Ethan Mollick getting blown away by Fable's capabilities, then Gary Marcus being a lot more cynical about…
Two Things Can Be True: Anthropic’s new Fable
I received two newsletter subscriptions within an hour of each other today regarding Anthropic's new Mythos-class AI model, named Fable. First, Ethan Mollick getting blown away by Fable's capabilities, then Gary Marcus being a lot more cynical about Anthropic's expert media manipulation. Ethan Mollick got Fable to do some truly exceptional stuff. And he writes really well, you can absolutely feel his awe coming through.
alanbuxton.wordpress.com
Two surprising things I heard at Snowflake Summit Joe Moura, CEO at Crew AI If it's a problem that the system gives a different answer each time, then maybe it shouldn't be an agent. This one cuts through a lot of the hype. Agents are amazing at many things, but it doesn't mean they are the right…
Two surprising things I heard at Snowflake Summit
Joe Moura, CEO at Crew AI If it's a problem that the system gives a different answer each time, then maybe it shouldn't be an agent. This one cuts through a lot of the hype. Agents are amazing at many things, but it doesn't mean they are the right tool for everything. Business systems need determinism (audit trails, controls, predictability, explainability etc).
alanbuxton.wordpress.com
AI vs Memento Working with Claude Code makes me feel like I'm in Memento, the 2000 film about someone trying to solve a mystery. The twist is that he can't make new memories. Every time he figures out something he has to resort to a patchwork of notes and tattoos so he has something to start from…
AI vs Memento
Working with Claude Code makes me feel like I'm in Memento, the 2000 film about someone trying to solve a mystery. The twist is that he can't make new memories. Every time he figures out something he has to resort to a patchwork of notes and tattoos so he has something to start from next time his short-term memory resets itself.
alanbuxton.wordpress.com
How do you benchmark a product you built yourself? I built a company-news API and I wanted to know whether it was better than the alternatives. The problem: I'm the author, so I'm biased — and I wanted to use an LLM as the judge, which makes it worse, because a model that recognises my product…
How do you benchmark a product you built yourself?
I built a company-news API and I wanted to know whether it was better than the alternatives. The problem: I'm the author, so I'm biased — and I wanted to use an LLM as the judge, which makes it worse, because a model that recognises my product (and works out it's being scored on behalf of its creator) has every incentive to soften the blow.
alanbuxton.wordpress.com
What makes a good data person? I found myself talking to two different data leaders last week in two different places. Two different leaders who work in completely different areas. Completely different. Yet they both used the same word to describe a great data person: "Curiosity". A curious data…
What makes a good data person?
I found myself talking to two different data leaders last week in two different places. Two different leaders who work in completely different areas. Completely different. Yet they both used the same word to describe a great data person: "Curiosity". A curious data person is someone who spends time in the data, finds interesting patterns and then uses that to create insights.
alanbuxton.wordpress.com
It’s a pause-calypse. AI is shifting the value from "renting software" to "owning intelligence." SaaS isn't dying, but the era of bloated subscriptions is; tools must now prove their worth by integrating deeply into custom AI workflows. Adapt or fade. — gen1e.xyz
SaaSpocalypse or SaaS-pause-calypse? The death of SaaS in the age of AI - e.g. people building and running their own CRM systems rather than using Salesforce - is overstated. I think we are more in a pause-calypse than an apocalypse1. The hype is overbearing. There is no end of people writing…
SaaSpocalypse or SaaS-pause-calypse?
The death of SaaS in the age of AI - e.g. people building and running their own CRM systems rather than using Salesforce - is overstated. I think we are more in a pause-calypse than an apocalypse1. The hype is overbearing. There is no end of people writing about the business they launched last weekend off a vibe-coded app.
alanbuxton.wordpress.com
Today's programming facepalm: Realizing I didn't have an index on a foreign key. Everything working a lot better now :D
AI uplifts your capabilities, but also forces you to uplift yourself This is a story based on some work I'm doing on my side project, syracuse.1145.am. It's an example about how with AI I can do much more than what I could do before and how I had to level up my thinking to make the most of it. The…
AI uplifts your capabilities, but also forces you to uplift yourself
This is a story based on some work I'm doing on my side project, syracuse.1145.am. It's an example about how with AI I can do much more than what I could do before and how I had to level up my thinking to make the most of it. The story is about python programming, and specifically about how scrapy manages its spiders.
alanbuxton.wordpress.com
What does “AI” even mean in 2026? Remember in the "BC" times (Before ChatGPT) when AI meant something different to what it does now? It's what we saw in science fiction films. It was something that either felt human-like or felt super-human. It was more likely to be logical than creative, but it…
What does “AI” even mean in 2026?
Remember in the "BC" times (Before ChatGPT) when AI meant something different to what it does now? It's what we saw in science fiction films. It was something that either felt human-like or felt super-human. It was more likely to be logical than creative, but it could still be wise. It was very hard to define it precisely. My favourite definition was…
alanbuxton.wordpress.com
UK Govt Entrepreneurs Call For Evidence is missing a critical question. Good news: The UK govt is looking for input into its strategy for encouraging more entrepeneurship. Link here in case any of my network have not yet submitted their views: It ends on 28th Feb so please do go ahead if you…
UK Govt Entrepreneurs Call For Evidence is missing a critical question.
Good news: The UK govt is looking for input into its strategy for encouraging more entrepeneurship. Link here in case any of my network have not yet submitted their views: It ends on 28th Feb so please do go ahead if you haven't already. Bad news: The questions highlight a blind spot at the heart of the UK establishment's view of entrepeneurship…
alanbuxton.wordpress.com
All this AI agent stuff going on and the real AI feature I would love is an AI that can tell me where my phone is.
Why the AI disruption feels different. Gen AI is changing everything. Gen AI is a bubble. Gen AI is coming for your job. The current AI disruption is getting more coverage than previous ones I've seen. It's because of the sorts of jobs that it is disrupting. It's affecting programmers. When social…
Why the AI disruption feels different.
Gen AI is changing everything. Gen AI is a bubble. Gen AI is coming for your job. The current AI disruption is getting more coverage than previous ones I've seen. It's because of the sorts of jobs that it is disrupting. It's affecting programmers. When social media started killing journalism there was a dark joke going round: "learn to code". Well, now the coders are among the ones being disrupted.
alanbuxton.wordpress.com
Ok I give up. I'll read the New Yorker article about if Claude is conscious, given that you lot are all quoting it.
After hours struggling with getting Microsoft Copilot to help convert some Loop pages to Loop components I gave up and asked ChatGPT. It helped me straight away. Go figure. TBH this isn't a "ChatGPT good, Copilot bad" post. It's "if one AI tool isn't doing it for you, try another one" post.
foreign [Music] Klangphonics recently turned up on my YouTube recommendations. Great musicianship, great sounds. Like the best bits of 1990's Trance. Love them. Here are some screenshots from one of their YouTube recordings with YouTube's auto-subtitles. At first, I found it entertaining how…
foreign [Music]
Klangphonics recently turned up on my YouTube recommendations. Great musicianship, great sounds. Like the best bits of 1990's Trance. Love them. Here are some screenshots from one of their YouTube recordings with YouTube's auto-subtitles. At first, I found it entertaining how YouTube struggled to subtitle this music. After all, this sort of music isn't really about the lyrics. I've no idea why YouTube says "foreign".
alanbuxton.wordpress.com
Wow! Syracuse beats Perplexity! My side project made almost a clean sweep of the AI analysis of the latest head-to-head between Syracuse, LinkUp and Perplexity. Grok, ChatGPT, Claude and Gemini all favoured Syracuse. CoPilot preferred LinkUp. Feeling quite proud because middle of last year there…
Wow! Syracuse beats Perplexity!
My side project made almost a clean sweep of the AI analysis of the latest head-to-head between Syracuse, LinkUp and Perplexity. Grok, ChatGPT, Claude and Gemini all favoured Syracuse. CoPilot preferred LinkUp. Feeling quite proud because middle of last year there wasn't much difference between one of these tools and another. To be fair, it was a very specific use case.
alanbuxton.wordpress.com