Peter Sergeant

Peter Sergeant

@sgnt.ai

Full-time building LLM and ML powered NPCs/agents for a popular online game. pete@sgnt.ai

153 followers219 following288 posts

Peter Sergeant
@sgnt.ai

Currently trying to get support from a company whose AI customer service agent does everything it can to dissuade you from talking to a human, and when you finally manage, simply passes you to a more advanced AI customer service agent who etc...

Peter Sergeant
@sgnt.ai

Both left and right can agree that Trump is terrible at hiring, the difference seems to be that the left think this is a skill issue, and the right seem to think that he's just consistently very, very unlucky, like he keeps getting struck by lightning or something.

Peter Sergeant
@sgnt.ai

Tired: SaaS vendors shoving AI into their apps Wired: adding assistants into your own personal-use vibe-coded tools

Peter Sergeant
@sgnt.ai

Whatever the opposite of “the benefit of the doubt is”, that’s the appropriate response to a Gemini release

Peter Sergeant
@sgnt.ai

I'm curious to see if we come up with designs to make it harder for humans to rubber-stamp LLM decisions. I run a system with LLM-assisted image moderation at the moment, and we had to have a word with one of the (volunteer, unpaid) moderators because they'd started to do that

Ethan Mollick@emollick.bsky.social

From The Lancet: in an urgent care setting, the advice of the obsolete Gemini 2.5 Pro & Gemini 2.5 Flash (without access to patient medical records) were rated of similar quality to doctors by other physicians. There was no safety issues spotted. Models have gotten significantly better since.

Peter Sergeant
@sgnt.ai

... even in the most egregious cases of the LLM getting it wrong. Alternatively, maybe this is a 2026-2027 problem that goes away as models improve. There's certainly a world in which the LLMs get good enough that all human intervention makes it worse.

Peter Sergeant
@sgnt.ai

I suspect MoD can think of better ways to spent £10b than on an anti-drone system for drones that somehow make their way from Russia across Polish, Danish, German, Swedish, Norwegian, and Dutch airspace unhindered before arriving in London. Tories are still not a serious opposition party.

Peter Sergeant
@sgnt.ai

Also incredible that the Tories keep discovering massive funding sources that they were unable to tap two years ago, after 14 years in power.

Peter Sergeant
@sgnt.ai

I understand we need an Opposition but we’re surely still a few years away from being able to take Tory statements seriously? Feels like we’re still in the “well why didn’t you?” phase of counter-proposals. If there were 23b savings in welfare to be found, how did you miss them before?

Peter Sergeant
@sgnt.ai

A couple of years ago I switched my job -- which had meetings -- for a job where I don't have meetings, but I get an endless stream of hate-mail on computer gamer forums. I think I'm reasonably happy with the trade tbh?

Peter Sergeant
@sgnt.ai

The Boy Who Cried Wolf is also a story about a village that loses a child to a preventable wolf attack due to cynicism. If the AI bros want us to take safety seriously, we should listen.

Peter Sergeant
@sgnt.ai

Somewhat convinced that the only way we're going to survive the first adversarial agent capable of finding and replicating its weights, is to kick one off now while the models are still comparatively dumb. Downside: definitely illegal and unlimited liability.

Peter Sergeant
@sgnt.ai

GLM5.3 with some stolen compute and excised guardrails feels like a good starting point. Hard to know who'd stick their neck out and set such a thing wild though.

Peter Sergeant
@sgnt.ai

Idea I’m fixating on recently, we’re going to hit a point quite soon (<12m) where for most humans, models are good enough, including free ones, and what’s more, humans aren’t really smart enough to notice the differences anyway 🧵

View full thread
Peter Sergeant
@sgnt.ai

This forces a significant rethink in the prop model + inference economics that OpenAI and Anthropic are relying on. Mass market falls out almost entirely and you’re left with people who need the absolute cutting edge. Research? Strategy work? Military? Hard to know 🧵

Peter Sergeant
@sgnt.ai

But the model becomes sell to people whose demand is price inelastic, because the general public are all just buying open-model inference on what have becomes essentially fungible tokens. END

Peter Sergeant
@sgnt.ai

Idea I’m fixating on recently, we’re going to hit a point quite soon (<12m) where for most humans, models are good enough, including free ones, and what’s more, humans aren’t really smart enough to notice the differences anyway 🧵

Peter Sergeant
@sgnt.ai

Like unless you’re a gamer, we’re already there for most people on how powerful their computer is. I think my phone is ~5 years old and my laptop is ~3 years, and I’m not upgrading because I’ll only notice a difference for a day or two 🧵

Peter Sergeant
@sgnt.ai

In 12 months, GLM and DeepSeek are going to be good enough for basically everything you need to get done. There will of course be differences, but presumably these will only come out head-to-head or for tasks that are already too advanced for most humans, like research 🧵

Peter Sergeant
@sgnt.ai

By Friday I’ve usually run out of Fable credits, so it’s a good time to try other models out. Running GLM 4.3 for security scans on several codebases and it’s doing a good job

Peter Sergeant
@sgnt.ai

I feel like the extraordinary claim that requires extraordinary proof is that this will somehow all blow over, not the claim that we're heading towards a singularity here.

Peter Sergeant
@sgnt.ai

Most of my AI experience at the moment (which agents to use, how to use them effectively, how to manage agent-driven projects) is coming from open-source side work, not my day job. Which is exactly the same way I got good at programming, twenty-five (or is it thirty?) years on.

Peter Sergeant
@sgnt.ai

Do Anthropic’s leadership realize how much developer goodwill they’re destroying by pairing their excellent model with the vibe-coded developer-anarchy bullshit that’s Claude Codex? I am so so close to switching my Max plan elsewhere

Peter Sergeant
@sgnt.ai

It feels like a national security issue that SecWar doesn’t understand that American military exceptionalism comes from being like Amazon, rather than Sparta

Peter Sergeant
@sgnt.ai

Fixing bugs with Claude is incredibly annoying, because if it detects even a whiff of security, is downgrades itself, and does a shittier fix.

Peter Sergeant
@sgnt.ai

My slightly insane workflow right now is when something is found, getting Opus 4.8 to reframe it and rewrite it to be purely a "containment / correctness" bug and drop all mention of security, start a new context, get Fable to fix it, and then come back to Opus (with auxiliary Codex and Grok checks)

Peter Sergeant
@sgnt.ai

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/...

wsj.comHarvard Has Trained So Many Chinese Communist Officials, They Call It Their ‘Party School’The university’s Kennedy School of Government has long been favored by party cadres seeking career boosts.
Peter Sergeant
@sgnt.ai

Cambridge (UK) has some of the country's best Chinese food and attracts some of the brightest minds because Xu Zhimo liked watching punts go by, and the Trump admin wants to screw up the Cambridge (USA) version of this

Peter Sergeant
@sgnt.ai

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

Peter Sergeant
@sgnt.ai

2. If I have a topic I want to brain-dump, I stick my headphones in and use voice mode, and ask it to pretend it's a Podcast host interviewing me ... only ask one question as a time, dive into anything interesting I just hint at. This helps me get over writers' (or programmers'!) block all the time.

Peter Sergeant
@sgnt.ai

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...

sgnt.aiUnderstanding Modern AI is Understanding Embeddings: A Guide for Non-Programmers (with lots of dogs!) | sgnt.aiEmbeddings are a core AI concept that underpin a great deal of what we today think of as being AI. This article is going to give you an accurate and intuitive understanding of what an “embedding” is i...