Building an AI that finds you leverage in the world... and knows when to say no. Happy to help with harness design and context engineering. 🇬🇧
UK
Joined May 2026
- Tweets99
- Following35
- Followers15
- Likes12
Pinned Tweet
Howdy, Twitter.
I made this account for a mix of reasons.
1. I'm cooking cool stuff and want to share it. I hear that building in public is what the cool kids do.
2. I've got a CS background, but my day job is as a creative professional and the old school creative vs tech bro discourse on generative AI is tiring and completely lacking in nuance on both sides. I hope to provide a pragmatic perspective that will upset both sides equally.
3. I want to meet people doing difficult things!
This starts with sharing progress on a couple things I'm building: and anti-sycophantic agent to help you find your personal edge in the world and a tool for tracking the probability of any event you care about occuring, without the polymarket risks.
Follow and say hi if that sounds interesting 👋
Claude Startup Program
Explain to me how Anthropic can award startups thousands of dollars worth of seats, vouchers etc... give them time to plan months of business activity on that... then revoke them completely?
And frame this as a 'pause' instead of a rugpull?
@AnthropicAI
99% of developers want as much inference as they can get with the best models
If OpenAI's plan is to drop out of that race and just wrap inferior models at inferior rates with software/bots, which you can build and host yourself using AI in an afternoon, they're finished #OpenAI
Asked 5.6 Sol High to perform a security review.
Came back 8 minutes later to find it had burned through 100% of my 5hr usage having achieved nothing.
I understand it's a more expensive model, but not 100x more costly - especially with the token efficiency gains?
Anyone else?
How impressed you are by frontier LLM's ability to act, plan, decide or reason is inversely correlated to how much you know about the domain in question.
'Building in public' is misunderstood.
Useful for feedback when it arrives, sure, but it has very little to do with marketing or distribution unless other builders are your market.
If you're selling software to dentists, showing your agent skill framework off on Twitter won't help.
Do it for community and fun. It will do absolutely nothing to sell your software.
It is sort of crazy to think of how much of the digital ecosystem, developed over decades and running the world, is rapidly collapsing into a handful of super apps that wrap LLMs.
Dan retweeted
Can see the intent, but prompts like this betray a lack of understanding of how LLMs work at a fundamental level.
You get wins from the behavioural influence towards specificity, patterns of thinking... absolutely.
But it boils down to 'be smart, don't make mistakes', which doesn't boost a model's efficacy or make it perform like a better model.
In fact, it spreads its attention mechanism over more things than the task you're giving it, which for long running or difficult things decreases its effectiveness.
Sycophantic LLMs will accelerate you in the direction you were already headed.
The more I use them, the more I want to write 'Your job is to win the counter argument' into the prompt of every single session codex or Claude session.
And still it's not enough...
With the massive growth of skill libraries and viral stacks that are just markdown files, I'd strongly suggest that people build and maintain their own.
The power of skills is their simplicity. Just a markdown file. By authoring them yourself in minutes, even if copying, you develop a critical understanding of them.
AI is already a black box, you're compounding that by installing endless skills you've never actually read #codex #claudecode