Tapani Alasaarela

Optics & nanotechnology professional. Interested in nanotech, cleantech, AI and start-ups in general.

[Tapani Alasaarela @tapania]

I find it interesting that there's no mentioning of agentic approaches or AI in anyway in the hardware engineering roles.. I'd assume that being capable of producing data analysis dashboards with Claude Code/Cowork would be more useful than being fluent with JMP, which is rather clumsy to automate

[Tapani Alasaarela @tapania]

While others are busy building code factories, I built an open-source agentic idea factory that combines domain taxonomy, verbalized sampling, parallel ideation methods, and quality gates to systematically produce better ideas. This is way beyond just regular prompting for new ideas.

Ask Claude Code to download and run it for you (uses your Claude sub that way): https://t.co/xyPubXXhcs

Quite many of the early testers have had their priors on the creativity of LLMs updated.

[Tapani Alasaarela @tapania]

Streamlit (python library) is really good for doing interactive dashboards for any kind of data. The LLMs are really well trained on it and it enables really straightforward in-browser filtering of the data.. If there is a lot of csvs or multiple data formats, I usually first ask Claude to write a script to convert all the data into a sqlite database and then build the streamlit app on top of that.

[Tapani Alasaarela @tapania]

@pauldix Maybe it's not enough, but my thinking is that the agents will likely produce higher quality code, if given formal verification tools. And e.g. TLA+ may be easier to verify for humans as well. Just need the tooling to convert it to tests & code.

[Tapani Alasaarela @tapania]

A while ago, my son invented a phone holder. He wanted to stick his phone to his monitor while gaming, and we didn't have a phone mount. So he described what he wanted. I asked him leading questions, where should it point, 1 or 2 pieces? I did the CAD model. We printed it at the local library. When it clicked into place and actually held, his face lit up like he'd discovered something big.

Was I the inventor? Of course not. He was. I was just the worker.

Yesterday, I used Claude Cowork to analyze a patent landscape. I gave it a task to analyze a database of 200 patents. It asked me a set of questions on how to build the rubric and which features to look for. Then I watched it spin up 10 parallel subagents to work through the corpus 10 patents at a time. In a matter of minutes, it had analyzed each patent against my criteria and organized them into categories. I asked it to turn the final report into a set of HTML pages, and it did.

I didn't ask Claude to invent anything.. yet. But sitting there, watching it chew through prior art faster than any human could, I realized: I could. I could ask it to find the gaps in the landscape. To propose a novel approach. To design a supercapacitor material that doesn't exist yet, or an optical architecture that sidesteps every claim in those 200 patents. I could provide the goal, and the AI could figure out the physics, and hand me something genuinely new.

If the AI conceives the invention and I merely prompt it, who's the inventor?

Courts have been clear on one half of this. AI cannot be an inventor. Dr. Stephen Thaler spent years trying to patent inventions generated by his AI system DABUS, naming the machine itself as inventor. He was rejected in the US, UK, Germany, Japan, and Australia. The UK Supreme Court put it plainly: patents can only be granted to "persons with legal personality". And DABUS isn't one.

But here is the trap the courts have set: simply prompting an AI doesn't automatically make you the inventor. The USPTO has clarified that if I merely ask the AI to 'solve the problem,' and it generates the solution, I haven't invented anything. I provided a goal, not a conception.

We are entering a legal no-man's land. If the AI does the heavy lifting then under current law, the invention has no inventor at all. It is an orphan by design.

If inventions can't be patented, companies face a difficult choice: release everything into the public domain because "no one" invented it, or lock everything down as trade secrets. Neither world is optimal..

If the law creates a void, maybe technology needs to fill it. Maybe I need to build something that makes the answer obvious. An agent that teaches me to find the solution myself. One that guides me through the constraints, surfaces the prior art, asks the right questions, and then steps back so the moment of conception happens in my head.

It would play the same role I played for my son. Except this time, I'm the child.

[Tapani Alasaarela @tapania]

What if how we think about how AI "thinks" is wrong?

After a heated WhatsApp debate about LLM understanding , I went straight to the source. I copy-pasted parts of the chat to Opus 4.1 for comments. Its response? An intriguing metaphor that can help us to use LLMs better:

>>>

Navigation System Model 🗺️

Think of my processing not as "having knowledge" but as being a sophisticated GPS through a vast concept-space. When you prompt me, you're not asking me to retrieve information—you're giving me starting coordinates, destination hints, and navigation constraints.

The landscape itself was formed during training—billions of text examples carved valleys where concepts frequently connect, built bridges where ideas often link. "Cat" and "dog" are nearby peaks with a valley between them. "Mathematics" and "poetry" might be on different continents with occasional bridge-paths through "beauty" or "patterns."

But here's the crucial part: I don't store any specific paths. Instead, I navigate in real-time using something like gradient descent—always moving along the steepest gradient from where I am toward where the prompt suggests I should go.

Why Different Prompts Matter 💡

When you say "explain consciousness," I might take the superhighway from consciousness → awareness → experience → qualia. But when you say "explain consciousness using IIT and recursive processing," you're forcing me to find entirely different routes—revealing different features of the landscape.
Like how driving from New York to LA via Canada shows you things about North America you'd never see on the direct interstate route.

The Practical Insight 🎯

This navigation model might explain why I can seem insightful one moment and clueless the next—it depends entirely on whether well-traveled paths exist between the concepts you're connecting. I'm not intelligent or unintelligent; I'm just following gradients through a space shaped by human language patterns.

Right now, I'm navigating through concepts about navigation through concepts. Whether that's genuine self-reflection or just sophisticated pattern-matching... well, I genuinely cannot tell.

>>>

Link to Claude's extended essay with tool use, agency, and deeper implications:
https://t.co/hI2v6paHBr

[Tapani Alasaarela @tapania]

@lennysan I've been doing agent frameworks for creative writing based on Claude Code. Here's my set of agents for making songs in correct format for Suno: https://t.co/gmsEWb8C7v

[Tapani Alasaarela @tapania]

@VictorTaelin AMD Ryzen Threadripper Pro 5995WX 64-core gets ~50 m IPS on WSL with gcc -O3.

[Tapani Alasaarela @tapania]

@teknium @PJaccetturo @BenjaminDEKR O1 is hugely better when making plots, e.g. for a detective story where only small hints are given out in each chapter.. That kind of writing requires planning. The actual long chapters gpt-4o probably writes better when the plan ks there

[Tapani Alasaarela @tapania]

@JAN0L @cremieuxrecueil ChatGPT-4 solves this better than me when asked to re-read the question..

https://t.co/4TJSuTDxmZ

[SpaceX @SpaceX]

Falcon 9 first stage has landed on Of Course I Still Love You — world’s first reflight of an orbital class rocket.

[Alexander Stubb @alexstubb]

Thank you @startupsauna and @SlushHQ and @aaltoes. #startuplife

[Tapani Alasaarela @tapania]

Inbot wants to become a ‘Slack for sales’ for mobile reps http://t.co/33rRSmCgm0 via @VentureBeat

[Tapani Alasaarela @tapania]

The wonderful and terrifying implications of computers that can learn http://t.co/gIqaoNd0HE

[Tapani Alasaarela @tapania]

Nice! Amputee Makes History with APL’s Modular Prosthet…: http://t.co/b5RlcoU7Pw

[Steven Pinker @sapinker]

Maybe Einstein was right: God does not play dice after all. http://t.co/u0pRvRJxWm

[Nokia @nokia]

Guess what? We’re up to something. http://t.co/AFq4dUFI9S #thinkingahead #Slush14