Building AI products. Ex. PM @Spotify | MLE @Spotify @BabylonHealth. Post about tech, mainly AI, and personal experiments. Sometimes wisdom from elsewhere
Joined September 2014
- Tweets457
- Following808
- Followers105
- Likes1.2K
Episode 8 is out! Steve and I go into brain-computer interfaces and all the interesting things that await us in our telepathic future
Listen on Spotify. open.spotify.com/show/092SOY…
@FrontiersPod
Worked on a toy implementation of GEPA, which is a neat approach to optimising text context e.g. prompts in your agentic system. github.com/Johnxjp/gepa-toy
If you sit and think about it, it’s absolutely magical what a language model can do. We’ve taught it how to code and with that whole worlds of digital creation have opened up.
John 🦄 retweeted
We’ve been experimenting with responsive generative video interfaces. Real-time video models are reshaping responsive design. A thread 🧵
John 🦄 retweeted
Real-time video models are unlocking new capabilities.
A first experiment: a drawing app that animates your sketches as you draw. Under the hood, a video model generates each frame live, turning every idea into motion instantly.
It costs to care about your work but I think it’s a price worth paying. Pride in your work supports the soul.
Slashed summarisation time hugely on Store by streaming transcript and llm response. AI meeting notes, still all local, now faster.
github.com/Johnxjp/store (macos)
I was putting way too much info in AGENTS.md. The models don't need it and it burns tokens.
Models are getting smarter and they need less hand holding. Much better to split things off or delete aggressively and only add back once becomes obviously necessary. Turn on Elon mode
John 🦄 retweeted
GPT-6 Astra represents a step-function change in model capability for interactive reasoning problems. It scores 66% on ARC-AGI-3 using our standard harness, and nearly 100% with a continuous conversation harness and custom compaction, at a cost of roughly $360 per game.
In fact, the continuous harness version significantly outperforms our human baseline in action efficiency across almost all levels. When we examined the reasoning chains to understand how the model operates, we found it performing highly efficient, on-the-fly symbolic world modeling for each game and level. It goes as far as developing its own shorthand DSL to represent in-game situations -- essentially a game-specific algebraic notation.
Overall, Astra exhibits symbolic modeling behaviors we had previously only seen with sophisticated harnesses -- so harness capabilities are increasingly shifting into the model itself.
We see Astra as a major breakthrough in model intelligence.
Read our post on Astra and what these results mean: arcprize.org/blog/astra
John 🦄 retweeted
Whoa! The best human Go player beat the best AI Go program (with a 2 stone lead). kedglobal.com/artificial-int…
The main point that stands out here is, if you can design the system with validation, an agent can keep working to solve it.
This is all possible due to a radically new method we’ve built to verify chip designs with AI.
We’ve developed a proprietary formal engine that increases verification quality autonomously, and enables AI reasoning and proving of properties about complex chip designs.
Using this engine, Redwood was fully verified without any human verification engineer. The first, and only, RTL drop from simulation to FPGA platform, had zero deployment bugs when running models end-to-end. Every block was taken to more than 95% code and functional coverage.
By getting designs correct near the first try, with very few iteration cycles, we reduce the reliance on commercial EDA tools that traditionally slow down chip design programs.
This will be the new scaling law for hardware design, where the verification capability scales with compute, not engineering hours.
John 🦄 retweeted
This is all possible due to a radically new method we’ve built to verify chip designs with AI.
We’ve developed a proprietary formal engine that increases verification quality autonomously, and enables AI reasoning and proving of properties about complex chip designs.
Using this engine, Redwood was fully verified without any human verification engineer. The first, and only, RTL drop from simulation to FPGA platform, had zero deployment bugs when running models end-to-end. Every block was taken to more than 95% code and functional coverage.
By getting designs correct near the first try, with very few iteration cycles, we reduce the reliance on commercial EDA tools that traditionally slow down chip design programs.
This will be the new scaling law for hardware design, where the verification capability scales with compute, not engineering hours.
"You're still solving problems with the wrong mindset"
Yesterday, I was in conversation with some top-tier AI engineers discussing agents and tool design.
I was surprised by their approach to solving problems.
In the past, it paid to be cautious and constrain default to deterministic patterns.
Models have become so good though that they now first ask can I solve this with an agent and a few basic tools.
It felt to me like quite a shift, and the main challenge these days is tool and harness design.
No one ever seriously speaks about ML Yearning from @AndrewYNg.
Okay it was unfinished, but I got more from that book about practical ML than I did from most of the other textbooks out there.
When I was at Babylon Health I used his guidance to answer that question that is a thorn in a side for every ML person "how much data do we need?".
He taught me to use learning curves to help estimate how much data would be needed to get to desired performance.
And, most of those fundamental ML concepts I think are still applicable today.
info.deeplearning.ai/machine…
A lot of people write agent prompts as a series of addendums e.g. don't do this and never do that. When that list gets too long it's likely a sign that you need to revise the decision guidelines you're giving the agent.
Granola IS for breakfast (sorry, marketing team).
I spent a couple of hours this week recreating Granola with Claude.
Why?
Well, they recently started charging to retain older notes and I felt I could just make my own.
And, given an agent spent a couple of days hacking its way through Hugging Face’s infrastructure, I figured cloning a meeting-notes app couldn’t be that hard.
A few hours with Fable over the week and I managed to pull together a decent clone. Plus, all the processing is local.
How?
To start, I just described how Granola worked from a user’s perspective and let it go. For the UI, I gave Claude screenshots and it nailed the layout. Like Granola, there’s no intrusive AI bot, audio is recorded in the background, transcribed and a summary is generated.
It’s amazing what is possible nowadays. I knew next to nothing about building a Mac app or processing audio but I knew that Granola records the system audio, so I gave Claude that hint. What’s more, when I noticed the transcript contained repeated phrases due to echoes, it taught me all about Audio Echo Cancellation and fixed it.
I will say, some technical knowledge did help speed things up. For example, I knew something about transcription models. Claude initially picked a slower model (Whisper) and I was able to point it in the direction of faster models (Parakeet) to compare. I had it build a quick eval to evaluate Word Error Rate on existing meeting notes so we could pick the best blend of speed and accuracy.
I was even able to hack together a decent system prompt by using Claude’s ‘/goal’ to automatically update the prompt until my summarised notes were close enough to Granola’s output.Â
It’s definitely not a perfect clone.
So what?
Sure it doesn’t have some of the sleek design elements of Granola, the speed of note generation is slower (although I’m sure I can improve that) and I didn’t finish it off with Chat and Templates.
But, I’m happy with it. And that’s the point.
It suits MY needs!
Personal software has never been easier to make.
What is still worth paying for then?
Coding agents certainly make it lightning quick to copy but features have always been copyable, just ask Snap — Stories got copied fast. Some things are inherently difficult to copy: data, proprietary models, etc. There are also softer assets like product judgement or peace of mind even.
Maybe Granola’s swish design is just the delightful cherry on top that makes it all worth it?
*What do you think?*
Try it out here: github.com/Johnxjp/store. A local macOS note taker.
Note, it takes up some RAM at the moment (details in the docs) but nothing a coding agent couldn’t help you modify.
