Putting AI to Work, Cloud1 and Vuo AI | Harri Puupponen, Pasi Jokinen | Neuvottelija 189

Summary

Harri Puupponen (Cloud1) and Pasi Jokinen (Vuo AI) came on to discuss how language models actually get into use — in companies and for individuals. The episode was recorded in April 2023, weeks after GPT-4 shipped and AutoGPT broke out, and the interview questions were written by GPT-4 itself.

The most usable single idea is perspective-shifting. Jokinen’s point is that a person cannot, on the fly, describe a complex or emotionally difficult matter from the other party’s point of view — working memory is not enough for the translation. Hand it to the model and ask for exactly that. Since the starting point of almost any negotiation is being able to state your own need in terms of the other side’s needs, this is a building block rather than a trick. His second tip is narrower and less known: ask the model for scientific algorithms for a problem rather than for frameworks or models, because that word produces the better list.

AutoGPT gets its clearest definition here: an executive function, the prefrontal-lobe counterpart to GPT. A loop takes a generic problem, has GPT break it into sub-problems, finds a solution pattern for each and expands outwards. The instance told to destroy humanity — ChaosGPT — ended up only tweeting about it. Jokinen’s interest is in the budget constraint: the run hammers a paid API, so the tool is given a budget and builds the best tree within it. Miettinen’s corollary is that when a complex query costs euro cents, a freemium model cannot be sustainable.

On hallucination, one of the guests recounts a conversation with a large Finnish bank whose lawyers had forbidden applications that hallucinate. The reply: the bank employs 13,000 hallucinating instances, and this particular hallucinating instance passes the California bar exam and doctoral-level finals. The honest boundary follows — at GPT-3.5 level nothing coming out can be trusted unless an expert validates it or the tool is used only for summarisation.

The best technical demonstration is where the model fails systematically. Take the river-crossing puzzle with the lion, the sheep and the cabbage, and change the rules so the lion eats both. GPT-4 solves it as the standard version anyway. Told something went wrong, it admits the error; asked to solve it again, it repeats the mistake; asked to list the permitted and forbidden combinations, it lists them correctly and still repeats the mistake. Data occurring at too strong a frequency makes escape from the hallucinatory bias impossible.

On whether an advantage can hold, Miettinen asks whether a Neuvottelija AI trained on 189 episodes would be steamrolled by a generic model within six months. Yes. Puupponen’s antidote is emotion and brand — the brand can be copied, the feeling it produces cannot. Jokinen’s is structural and is the episode’s most original argument: the village smith was interchangeable because the role was what mattered, which is why the world is full of Smiths; now the abstraction level has risen, expertise has millions of niches and each has its best practitioner. People want Jony Ive, not “a designer”. Language models build very strong symbolic representations around such names, so products will keep being sold under professionals’ names — a Jony Ive design agent inside an AutoGPT run, a Sami GPT to negotiate the deals.

On M&A, Miettinen relays Bengt Holmström’s observation that full disclosure in a data room produces informational equality, which makes deception impossible and yields the perfect price — set against the Trojan-horse risk of a competitor taking everything and withdrawing. Jokinen’s middle path is machine-run due diligence against a trusted third party’s validated open DD model, returning only pass flags until a firmer offer justifies full transparency. Puupponen closes the section with the limit: AI cannot create trust; it is built from small words and unsaid moments.

The episode ends on hardware. Cost-effective applications have multiplied perhaps a thousandfold, so computation explodes. At the time of recording GPT-4 had rate limits and was available in Azure only from the United States — “we are being sold a bit of scarcity”. NVIDIA’s GPU forecasts are absurd and, Jokinen argues, still insufficient; demand will exceed availability for running and training large models, and the question becomes who can afford the capacity.

A recurring thread worth noting: both guests independently arrive at organisational shape as the binding constraint on what the technology will actually do. Jokinen cites, without being able to name it, the law that complex systems end up mirroring the communication structures of the organisation that built them — Conway’s law — and Puupponen confirms it from delivery work, where however sensible the technically best answer, the result ends up shaped like the customer’s org chart.

A note on the source

The transcription contains one gap of roughly 27.7 seconds at 00:55:40, in the middle of the WebGL demonstrations: speech breaks off after the Tower of Hanoi request and resumes at the tamagotchi request. This summary is written only from what the transcript contains; the gap has not been filled by inference. The MacWhisper transcript carries no speaker labels, so speakers are named only where the transcript gives direct evidence — being addressed by name, or a self-reference only one guest could make. Two passages whose speaker could not be established are attributed to “one of the guests”.

Watch

The recording lives on the Neuvottelija channel: Tekoäly käyttöön Cloud1 Vuo AI | Harri Puupponen Pasi Jokinen | Neuvottelija 189. A Finnish edition of this episode is published at www.neuvottelija.fi.