The intelligence explosion has only just started. The market goes to whoever uses it best.
I'm Vilhelm Toivonen, a technical AI partner. I help your company get more out of AI than your competitors do.
Book a 30-minute call with meIn 2022, this job needed a data centre.
An answer as good as the original ChatGPT's got more than 280 times cheaper in two years.
Now it runs on your device. No cloud, no cost. Try it.
In 2022, this job needed a data centre.
An answer as good as the original ChatGPT's got more than 280 times cheaper in two years.
Now it runs on your device. No cloud, no cost. Try it.
Kimmo: the switchboard for the school is delayed again, now week 47 Kimmo: Aleksi can you tell the main contractor Pekka today? Aleksi: ok. the cable trays for floor 2 are missing too Kimmo: Laura pls order the trays this week. and decide if we start wiring the other floors while we wait
Works best with one conversation at a time.
The model is small, reads English only and is tuned for this kind of text.
Approve the tasks once they look right to you.
Running on: your device · Your text sent: 0 bytes · Cost: €0.00 · 0 tokens/s
The model runs on your device. What you type into it stays there.
In 2022, this job needed a data centre.
An answer as good as the original ChatGPT's got more than 280 times cheaper in two years.
Now it runs on your device. No cloud, no cost. Try it.
In 2022, this job needed a data centre.
An answer as good as the original ChatGPT's got more than 280 times cheaper in two years.
Now it runs on your device. No cloud, no cost. Try it.
Kimmo: the switchboard for the school is delayed again, now week 47 Kimmo: Aleksi can you tell the main contractor Pekka today? Aleksi: ok. the cable trays for floor 2 are missing too Kimmo: Laura pls order the trays this week. and decide if we start wiring the other floors while we wait
Works best with one conversation at a time.
The model is small, reads English only and is tuned for this kind of text.
Approve the tasks once they look right to you.
Running on: your device · Your text sent: 0 bytes · Cost: €0.00 · 0 tokens/s
The model runs on your device. What you type into it stays there.
The model did the work in 12.3 seconds. You took 104.5 seconds to approve it.
The bottleneck was you.
You were quick. In a company, the approver is rarely there right away.
Inside companies this happens every day now.
Want to talk? I answer personally within two working days.
The work got faster. Why didn't the processes?
The processes were built last millennium, for slow work and for trust between people. Some steps are now many times faster, but you still finish only as much as the slowest step allows.
Example: an agent finds a security hole and writes the fix in minutes. Fixing has become many times faster, and so, unfortunately, has attacking. Still, the fix doesn't reach production until a person has reviewed it.
Example: an agent finds a security hole and writes the fix in minutes. Fixing has become many times faster, and so, unfortunately, has attacking. Still, the fix doesn't reach production until a person has reviewed it.
In the future version, the AI doesn't need to get any faster.
The gain comes from what the AI is trusted to do safely without a separate human check.
In the future version, the AI doesn't need to get any faster. The gain comes from what the AI is trusted to do safely without a separate human check.
AI can stall a process in two ways.
The model is too weak, or the process won't let it do anything. Many people still know AI from free ChatGPT or Copilot, which answers questions and does nothing else, and they build their processes around a weak model.
Today the limit is usually the process. You can tell because a new frontier model no longer changes the results. There are too few integrations, the AI gets no authority, nobody follows its work, and nobody knows what it can already be trusted with.
Models get smarter and cheaper every year.In a good strategy, the process improves by itself whenever a new model comes out. When the AI sees your error reports and customer feedback, every new model makes your product more reliable.
Build your processes around your best people. An electrical designer who handles two projects at a time now will handle four as the models improve, then ten, then twenty. Their domain knowledge and their own way of designing make each project better than the AI would on its own. The best programmers are still the best with AI. They just finish many times more.
The other half is the technology.
An AI product is only as good as its weakest layer.
- Infrastructure engineer
- AI researcher
- AI engineer
- Security specialist
- Product owner
- Vilhelm
In my experience, few people in Finland properly understand even one of these layers.
I've worked in every one of them. I've trained models, researched inference for my PhD and taken agents into production for hundreds of thousands of users.
What I've built.
Head of AI at Bondata, a Finnish employee and market analytics group. Previously CTO of Vibemetrics.
Research programmes.
I've designed and led large AI research programmes at Bondata.
Agents in production.
I've built scalable agent systems used by hundreds of thousands of people.
Foundational research.
I've done foundational research on efficient AI at the University of Helsinki.
Startups.
I've found laborious processes and replaced them with AI systems that need a fraction of the resources.
Two ways to work with me
Focused project.
In two weeks you know what in your company stays valuable as AI improves, which processes to rebuild and in what order. The plan is built to get better with every new model.
Long-term partnership.
For companies where AI becomes core. Models improve every year, so the processes have to improve too. We build them together and keep them current.
In half an hour we can look at which of your processes would gain most from AI.
