Forget the complexity: AI all boils down to drawing the right lines: sgnt.ai/p/finding-li...

In RAG, your chunks solve three distinct problems; trying to use one algorithm to solve all three probably isn't optimal: sgnt.ai/p/rag-trinity/

One day in 2012 I wrote up from a nightmare in which I was being chased by angle brackets, and sweating profusely, and found this code in my editor github.com/pjlsergeant/...

I will fix your vibe-coded MVP. Not joking. sgnt.ai/p/vibe-coded/

China sends its most promising civil servants to a US university for 1-3 years, and rather seeing this as a total and complete Cultural Victory, recruitment opportunity, and chance to indoctrinate them, the Trump administration wants to shut this down. Baffling. www.wsj.com/world/china/...

Three dumb ChatGPT tricks I use all the time... 1. If I need work critiqued, I tell the LLM that I got someone else to write it for me, and I'm going to pay them based on how good it is. Less time spent softening feedback, and more time spent on pros and cons. Repeat we get to a solid 9 or 10/10

I finally completed an article I've been trying to write for a year; the non-programmer's guide to embeddings! No maths, lots of dogs, and a step by step intuitive guide: sgnt.ai/p/embeddings...

It's final paper season.

Let’s say we were to discover tomorrow that dolphins are in fact fish, not mammals. Given the huge amount of data in LLM training sets asserting that dolphins are mammals, how does that fact end up getting updated? Just that we generate more new text as time goes on and we eventually have more …

Our LLM game NPC couldn't block chat input & got interrupted constantly ... we engineered "killable" responses using cancel tokens + input debouncing so it could pivot instantly like a human. Made interactions way more natural for impatient players. sgnt.ai/p/interrupti...

In-memory free-text search is a super-power for LLMs; spend less on inference by sprinkling in some free lexical search into the prompt… sgnt.ai/p/free-text-...

From OpenAI’s agent building advice. The first two of these are terrible, terrible ideas for things to give to a LLM to make a decision on. More: sgnt.ai/p/hell-out-o...

I would love to see more research about whatever we’re calling Sapir-Whorf for LLMs. Do you get substantially different opinions from an LLM depending on which language you query it in? Does its opinion on the greatest empire of all time change depending on if your query is Mongolian or Italian?

Don't let an LLM make decisions or implement business logic, they suck at that: sgnt.ai/p/hell-out-o...

Four bad definitions of agentic AI: sgnt.ai/p/agentic-ai...

Is anyone else building behavioural plugins for agents / chat-bots? Certain modes can be enabled for brief periods that add new functionality depending on triggers? I keep having to invent novel stuff for this project, and it would be amazing to chat to people doing anything vaguely similar

Remake of Eternal Sunshine of the Spotless Mind, only this time the guy’s seducing her based on her ChatGPT history

Please can we not radicalize Mark Cuban by being mean to him online

hold up

If they've RLHF'd out the ability for ChatGPT to check its own working I'm gonna be pissed

I’m not going to link to the article because we don’t reward clickbait in this house, but claiming that “generative AI is a con” because OpenAI and Anthropic haven’t proven that prop-model plus selling inference is a good business model is dumb

Full marks for self-confidence but _nil points_ for execution there buddy

I’m aware of two reasonably good definitions of Agentic AI, but I can definitively tell you that in practice it just means “stuff that AI might be able to do for us some day”. I haven’t heard anyone use it as anything other as a stand-in for “awesome stuff coming soon” in 3 days of this expo

I’ve done a few decades of software engineering and never had to deal with a field before where keeping the ability to rip out components is such an advantage as AI

At an AI expo in Dubai. Nothing destroys an exhibitor’s credibility faster than having the word blockchain anywhere near their product

One thing academics should take away from Deep Research is that a substantial number of your readers in the future will likely be AI agents. Is your paper available in an open repository? Are any charts and graphs described well in the text? Is it well-titled? Probably worth considering these…

Blog article to come when I'm not insanely busy, but had good results today with adding a "read the room" instruction to a the CoT prompt. The problem I'm trying to solve is the user telling the bot "my favourite sword is the Katana!" and the bot just providing info the user doesn't need about it

I mean, I'm not saying anything particularly novel here, but DeepSeek is really very very good, and I'm not sure how ChatGPT stays as my daily driver at this rate

Writing the perfect non-technical article about embeddings is haunting me. I have rewritten this damn thing four or five times, each time getting stuck on a given topic transition being too stark. It’s so close I can feel it, but I’m sure I’ll still be saying that in a year

"Street-fighting RAG: chain-of-thought prompting", in which I start talking about the challenges of developing a RAG application in what I am utterly convinced is the most challenging possible environment... www.sgnt.ai/p/street-fig...

Many thousands of messages to analyse in the New Year spewed forth from our intentionally very abusive and mean NPC. While he's exceptionally crotchety, angry, violent, and sweary, it's also important he's not racist, sexist, or ableist. If this would be useful synthetic data for anyone, do lmk

I’m at a Christmas Eve party and the players found the first NPC. Thousands of messages in the last hours, and here I am watching it happen on Langfuse from my iPhone 😅 Send hug-ops

Langfuse's possibilities and web interface are sparking joy, but Langfuse's JS/TS documentation is absolutely _not_

Not my strongest effort, but a pretty coherent Scrooge NPC character in production after a couple of days of effort (building on a much larger body of work I'm putting together). He's, uh, a little angry. Bah humbug!

I created a Character AI account to see if you can interrupt characters while talking (you can't). I sent maybe four messages to a random bot. It's been emailing me every day since. Perhaps I am meant to cure _it's_ loneliness?

What's SOTA for contextual chunking of PDFs these days? I have a reasonably clear picture of how I could whip something together that I think would be pretty effective, but is there something recent, excellent, and obvious?

I feel like the technique described could be naïvely implemented as “generate the output a few times with some reflection and then get the bot to summarize what it thinks are the best answers”

AI has _not_ been able to convincingly recommend other songs to go on my playlist that leads with Wayfaring Pilgrim by Roy Buchanan and Al Kooper playing Season of the Witch. Can Bluesky do any better?

One of the advantages of developing for a game is we can actually deploy AI agents -- obviously there is endless testing, evals, etc, but if our LLM-backed negotiation bot accidentally under-prices an asset it's not the end of the world. The next year is going to be incredibly exciting

I’ve completely stopped Twitter use for … a month now? Turns out I just wanted something to scroll and there’s now enough here to scratch that itch

Other than the PleIAs models, are there any other truly open-source and not-copyright-encumbered models?

It's _almost_ too obvious to need saying, but your life will be 50x better if you can very, very quickly test prompts (without running the whole eval suite) and then have all the logs you might possibly need at your fingerprints

One of my favourite things to do recently with ChatGPT is voice-mode to ask it to interview me about something I need to write. Specify I only want one question as a time, and then dialogue with it. At the end, it can summarise and organise what I said

Am I hallucinating this, or are none of the "fast inference" companies actually ready to sell yet? Groq has pricing, but you can't sign up for billing. Cerebras has no pricing (and a tiny context window). SambaNova, no pricing.

Gemini 2.0 flash: speed looks good vis-a-vis 4o-mini, and quality looks good so far against my eval set. If it's cheaper than 4o-mini too (which, it probably will be?) then OpenAI have a real problem, because switching between them is a value in a config file.

All I want for Xmas is legislation that bans apps and sites from disabling paste in password fields

Have any of you actually made generating SPARQL queries against a knowledge graph work in practice on a real deployed project? I’m pretty skeptical of how feasible this is in the real world

Is having a truly open model create lots of synthetic tokens, and then using a non-free model decide which of them are suitable for further training, copyright laundering?

When, if ever, do QAs / traditional test developers take over writing the LLM evals? I have a strong interest in testing, and hearing almost nothing from that part of the wider dev community

A test of how seriously your firm is taking AI: when o-1 (& the new Gemini model) came out this week, were there assigned folks who immediately ran the model through your internal, validated, firm-specific benchmarks to see how useful it as? Did you update any plans or goals as a result?