EP189 · Tools · first published 2023-04-27
Putting AI to Work, Cloud1 and Vuo AI | Harri Puupponen, Pasi Jokinen | Negotiator 189
Harri Puupponen, founder of Cloud1, and Pasi Jokinen of Vuo AI work through how companies actually put language models to work. Recorded in April 2023, weeks after GPT-4 shipped and AutoGPT broke out, with the interview questions written by GPT-4. It contains the single most usable trick for a negotiator — perspective-shifting — plus AutoGPT defined as a prefrontal-lobe counterpart to GPT, the anecdote about a bank that employs 13,000 hallucinating instances, a demonstration of why GPT-4 cannot escape over-represented training data, Bengt Holmström on informational equality in due diligence, and a forecast for when GPU capacity runs out. Published 27 April 2023.
Putting AI to Work | Harri Puupponen, Pasi Jokinen
Summary: In episode 189 Sami Miettinen interviews Harri Puupponen, founder of Cloud1, and Pasi Jokinen of Vuo AI about how Finnish companies and individuals actually put language models to work. The episode was recorded in April 2023, a few weeks after the release of GPT-4 and the sudden rise of AutoGPT. Published 27 April 2023.
The setup: GPT-4 wrote the questions
The episode opens unusually. The two guests had been running a long WhatsApp thread about AI between themselves, and GPT-4 was asked to write two questions for each of them as the basis for the conversation. The questions are read out verbatim and answered. The setup is itself the first argument: the model produces questions that survive being read aloud in front of experts. Jokinen notes that GPT-4’s question drew on his own “Sami GPT” — material about Miettinen fed into the model.
Vuo AI: a replacement for the non-fiction book
Pasi Jokinen states Vuo AI’s goal plainly: to build, with AI, a replacement for the non-fiction book. The reasoning is experiential. Anyone who has opened ChatGPT has noticed they can pull deep content out of any subject faster than by reading a book, and from that comes the intuition that the future of transmitting knowledge, learning and applying theoretical models lies in this direction.
The company has deliberately chosen not to build the model itself. Jokinen names the fear that drives the whole field: whatever you start building, within three months the technology can leapfrog you and make your work unnecessary. So Vuo AI looks for cornerstones that survive the churn underneath. The most important is the interface — the human interface and the interface to other machines. In practice that means a prototyping tool for rapidly testing complex prompts and prompt combinations and then using them programmatically.
He frames the goal through the service designer’s role: where a service designer was once relegated to describing what the customer wants and passing it on, the aim now is to give them the tools to play a central part in creating the software’s logic — through prompts. The most concrete example is the episode’s funniest: his son built a tool that takes a child’s request and rewrites it in a form his mother finds easy to understand and easy to say yes to. The first application was raising his pocket money to ten euros.
Cloud1: a data house that gained a new business line from OpenAI
Harri Puupponen comes at the subject from the opposite direction, and somewhat reluctantly. At Cloud1, machine learning has been ordinary project work around data and analytics for years, and among data people “AI” has been a word that draws frowns: the debate is too popular, and once you get to the technical facts you always discover it isn’t real AI after all. With ChatGPT and OpenAI, he says, he has given himself permission to talk about artificial intelligence.
His commercial point is precise. Cloud1 works with Finland’s largest companies on data, and the most interesting thing about OpenAI is the ability to build a customer their own ChatGPT on their own company data, which they do not want released into any uncontrolled state. This is one of Cloud1’s future business areas.
Miettinen asks bluntly why a “slightly boring Azure cloud company” is suddenly interesting — the background being Voland Partners’ €40 million fund and a Kauppalehti report of interest in Cloud1. Puupponen’s answer is a defence of the narrow patch: the company picked the data angle within cloud specifically and predicted that cloud and data would bring growth. That is what happened, and demand is now growing faster than the market can supply.
The practical negotiation edge: shifting perspective
The episode’s most immediately useful passage is Jokinen’s description of what an individual can do right now. He calls it perspective-shifting. If you have a complex or emotionally difficult thought in your head, you cannot describe it from the other party’s point of view on the fly — working memory simply is not enough for the translation. Throw it at ChatGPT and ask it to describe the matter from the other person’s perspective, and the result is remarkable.
From the standpoint of Miettinen’s book Uusi neuvotteluvalta (WSOY) this goes to the core: the starting point of a negotiation is almost always being able to describe your own need from the other side’s perspective and in terms of their needs. It is the basic building block everything else rests on.
The second practical tip is technical and less well known. Do not ask the model for “frameworks” or “models” but for scientific algorithms for solving a particular problem — in Jokinen’s experience the word algorithm produces the best results. You get a list of the best models, pick one, and ask the model to ask you for the contextual details it needs to run that model against your case.
Closed hierarchy, market or network
Miettinen brings in a conversation with futurist Risto Linturi about organisational theory: in some situations it pays to do things yourself inside a closed hierarchy, because transaction costs disappear and you can direct work by managerial authority without paying a market price; in others the problem is better pushed out to the market or a partner network. The question is what AI does to those interfaces.
Puupponen’s answer is that inside a hierarchy a model can be enormously effective, but not everything has to be paid for if you ask nicely — and if AI can be harnessed to ask nicely, the cost level falls. Jokinen extends this into the concept of a trust account: people in organisations also trade non-monetary favours according to whom they trust. His analogue is the way countries form multilateral blocs where cooperation is not transactional money but a circle of trust. AI can create means of increasing that trust and replicating it.
Puupponen’s own example is reading a non-fiction book with ChatGPT: he asks the model to summarise the book chapter by chapter with the main points and the essential examples, because he cannot motivate himself to wade through 200 pages for the handful of things that actually stay with him. Miettinen notes this leads straight into a copyright question.
AutoGPT — the AI of AIs
Jokinen introduces AutoGPT, which at the time of recording was barely a month old and had gained popularity on GitHub faster than any open-source project before it. His definition is his single best formulation: AutoGPT is an executive function, the prefrontal-lobe counterpart to GPT. It has a loop that drives problem-solving. You give it a generic problem, it asks GPT to break it into sub-problems, leans on Google to find a solution pattern for each, and builds an ever-expanding network of how the main problem gets solved.
The limits showed up quickly. When one instance was told to destroy humanity — ChaosGPT — it ended up only tweeting about destroying humanity. The detail Jokinen finds most interesting, though, is the budget: because the run hammers a paid API and GPT-4 is not cheap, you give the tool a budget constraint and it builds the best possible solution tree within it. Miettinen picks this up: when a complex query costs euro cents, a freemium model cannot be sustainable.
Jokinen describes where this leads through his own experience. He had written a five-step list for getting sufficient understanding of any abstract problem quickly and building a tool for a professional. The list was run by hand as five consecutive GPT-3 calls. When GPT-4 arrived he threw all five steps in at once — and the model simulated the whole execution in a single pass. He calls this a rise of two abstraction levels: the model does not merely perform five tasks in sequence, it rises above the thought in a way that “feels like consciousness, even though I know it isn’t”.
Hallucination: “you employ 13,000 hallucinating instances”
Miettinen asks whether hallucination is a fair term. Both guests find it usable — it cannot be called creativity, and it points in roughly the right direction even if it does not land exactly.
Then comes the episode’s most quotable anecdote. One of the guests describes a conversation with a large Finnish bank whose lawyers had imposed a constraint: no applications that hallucinate. The reply was that the bank employs 13,000 hallucinating instances — and that this particular hallucinating instance passes the California bar exam and doctoral-level final exams. Do you think of your people as error-free?
What follows is an honest boundary. At GPT-3 and 3.5 level you cannot trust anything that comes out: either an expert validates it, or the tool is used only for summarisation, where the source data comes from elsewhere. With GPT-4 it starts to be the case that some things can perhaps be trusted — and connecting a search function brings curated data and real-time information into the mix.
Where the model is genuinely bad: the lion, the sheep and the cabbage
The episode’s best single technical demonstration concerns systematic failure. If a dataset appears very frequently in the training material in a completely fixed form with one “correct” solution, GPT-4 is almost incapable of getting past it.
The example is the classic river puzzle: a lion, a sheep and a cabbage on the bank, one at a time in the boat, and the wrong pair left ashore means one eats the other. (Miettinen notes the same puzzle featured in Reaktor’s AI course as a decision-tree exercise.) When the rules are changed — the lion eats the sheep and the lion eats the cabbage — the model still solves it as if it were the standard version. Ask whether something went wrong and it admits the error. Ask it to solve it again and it makes the same mistake. Ask it to list all permitted and forbidden combinations and it lists them correctly — and still makes the same mistake.
The conclusion is one of the episode’s sharpest lines: data occurring at too strong a frequency makes it impossible for the model to escape the hallucinatory bias. What the model ultimately does is predict the next most probable word.
Does the advantage hold? From the smith to Jony Ive
Miettinen puts the depressing question: if you built a Neuvottelija AI and fed it the scripts of 189 episodes plus the community’s accumulated wisdom, would a generic model steamroll it within six months? The answer is short: yes.
Puupponen takes this into ownership. Copying a good business idea is extremely easy, and if the copy is similar enough but different enough to sidestep the law, you end up in a world where it is hard to say who owns what — whether abstract things can be owned at all. His antidote is emotion and brand: you can copy the brand, but not the feeling the brand produces.
Jokinen’s answer is more structural and is the episode’s most original thought. Centuries ago the village had a smith, and it made no difference whether the smith died and the apprentice took over, because the role was what mattered — which is why the world is still full of Smiths. Now the abstraction level has risen: we have unique names, there are millions of areas of expertise, and each has its best practitioner. People no longer want “a designer” — they want Jony Ive. This aligns with how language models represent the world: very strong symbolic representations form around labels with highly specific meaning. So he believes that even ten years out, products will be sold under professionals’ names: an AutoGPT product design run brings in a Jony Ive designer agent, and “Sami GPT will negotiate my deals, because I know it’s very good at it”. Expertise does not disappear; it becomes callable.
The validation network: a negotiator’s vision
Jokinen sums up his own picture of the future through AutoGPT’s weakness. The tool creates the problem-solving model on its own, but every proposed solution is deficient in some way — they need validating. In his vision, when a new business opportunity appears, someone at the top writes the master query, ten high-level sub-problems fall out of it, and each one automatically sends a validation request to the relevant top expert in the network: high level, is this fine? A confirmation comes back within three seconds, and each validated component spawns ten further steps beneath which the same thing repeats. The result is a recipe for an extremely complex problem, produced in minutes or hours and validated by the network — the design of a system the size of Finland’s Apotti healthcare platform, say.
Puupponen supplies the missing piece: negotiation carries a great deal of emotion and interpretation, down to micro-expressions. Miettinen says this has been practised by filming negotiations and analysing micro-expressions — and that he either controls his emotions well or is a psychopath, whereas the other participant was in a continuous state of fear. Puupponen’s follow-up question is the essential one: when the structured approach is available to everyone via AutoGPT, the differentiator is who holds the data on micro-expressions — and soon nobody will want to release data like that publicly, because it ends up generically in the AI maw.
Trust, due diligence and Holmström’s informational equality
On the M&A side Miettinen returns to his interview with Nobel laureate Bengt Holmström. Holmström’s observation is informational: in an auction process the investment bank makes a hundred contacts, which refine through info memos shared under NDA into dozens, from those a handful of offers that receive a management presentation and more data, and finally, under exclusivity in the data room, all available information is disclosed. It is in the seller’s interest to open up completely, because that is what achieves informational equality, which makes deception impossible and produces the perfect price.
The risk is obvious: if the counterparty turns out to be a Trojan horse — a competitor in the same industry that takes all the information and withdraws citing strategy — the damage is extreme, legal sanctions notwithstanding. Afterwards that party is blacklisted in the investment bank’s books, and Miettinen asks whether such information ought to be shared more widely.
Jokinen’s middle-ground solution is the episode’s most concrete AI application to M&A: if a trusted third party — an investment bank, say — has validated an open DD model, machine learning models can run the due diligence and return only flags: passes, passes, passes. Only once everything passes and a firmer offer is on the table does the process move to fully transparent DD. The bidder gets near-complete validation and can commit more. Miettinen recognises this as a variation on the clean room used when two competitors cannot share market information with each other but a trusted third party may hold it.
One of the guests recounts a cross-border M&A process in which a buyer was sought globally and a second international investment bank was engaged to maintain communication between the bidder and the seller. The essential observation was that neither party wants to reveal anything — least of all anything to do with price — and that the greatest value in the whole process came from how the party oiling the communication could, through constant contact, draw a little more information out of each side and reveal it to the other in a way that felt safe. From this he concludes that every relationship needs an investment bank — an outside party with a motive to make the cooperation work. Miettinen calls it a consigliere solution and adds that the psychology of expectation management is central to it.
Puupponen closes the section with what AI does not do: AI cannot create trust. Trust is built out of the small words and the unsaid moments of communication, and the few people who can create it where there is no ground for it have a skill whose value only grows.
Playbooks, Conway’s law and the shape of the organisation
Miettinen describes asking Blackstone’s global chief executive, at an invitation-only event, how the firm’s playbooks have developed with AI. A playbook is the general operating instruction applied across hundreds of private equity holdings: if inflation rises, what is the best practice all portfolio companies should adopt? The answer was that development is rapid, but the silos are partly separate and company-specific information cannot be shared — it is precisely the generic operating models from which, with anonymised data, genuinely good recipes can be drawn.
Jokinen equates the playbook with AutoGPT’s solution hierarchy and predicts that products will start to be built as AutoGPT-type partial solutions that are business wholes, with a new abstraction level above them — tied to the financial market, for instance — affecting everything below in the manner of a playbook. If the recipe is sufficient to create an entire information system, then when a significant change occurs you alter the underlying instruction set and run the creation process through again.
Here he refers to a law whose name he cannot recall: all complex systems end up following the internal communication structures and hierarchy of the company that created them. (It is Conway’s law, which neither guest names in the episode.) If Microsoft has three groups with their own responsibilities, the product will contain three distinct components. Puupponen confirms this from his own work: however sensible the things technology could do, the use of technology reflects how people and the business are arranged in the organisation — and the result looks like that arrangement. His conclusion is that the structure persists but the people are gradually replaced: low-level work disappears and the system’s role grows relative to the humans’.
Office, Power BI and WebGL — everyday productivity
Miettinen predicts that Office will gain auto-prompting and suggestion: email replies moving from three stupid options straight to text. Puupponen confirms the direction via Copilot and laughs that he “codes in PowerPoint” — since his own decks resemble one another, a tool can generate the rest from the heading. The return of the paperclip is predicted half in jest. The substantive point is that the organisation’s entire body of data, from Word templates to PowerPoint, becomes available, so mechanical work speeds up and that hits labour productivity immediately.
Power BI gets its own passage. Puupponen says it was brought in as an extension of Excel for data analytics, for people who think in SQL, and that this is exactly the area AI runs right over: ChatGPT helps a great deal with building Power BI, can write basic SQL, simplify complicated code, and explain what to do to get data into the shape you want.
Jokinen illustrates coding ability with two experiments. His first request to GPT-4 was to draw a WebGL scene with a white plane, a curved line and a rainbow-coloured ball travelling from start to finish — and the model produced it, animated in a Three.js environment. The second request was made in Finnish: animate a solution to the Tower of Hanoi problem. A third was “write me a virtual tamagotchi in a programming language of my choosing”, and the foundations came back immediately. Puupponen adds the cloud person’s next step: make a cloud-suitable version of it that can be abstracted away somewhere to run.
Out of this comes the episode’s finest single insight about productivity. TikTok and fast-moving devices have ruined human concentration, but GPT technology has unlimited concentration. It is therefore the perfect complement to a human whose attention horizon is zero: it gives back the concentration we have taken away from ourselves. It is enough to concentrate for fifteen seconds, write, throw it out, come back — and it was done.
The hardware ceiling: GPUs, Azure scarcity and the edge
The constraint at the end of the episode is physical. Jokinen’s claim is that the most important limit on using language models is simply that hardware capacity runs out. The number of applications that can be built cost-effectively has grown perhaps a thousandfold or a millionfold with language models, which means the amount of computation explodes. The question becomes whether you can buy enough OpenAI API instances for next year.
Puupponen compares this to the start of the pandemic, when Teams usage spread and cloud capacity ran out. At the time of recording GPT-4 has rate limits in place and is available in Azure only from the United States, not Europe — “we are being sold a bit of scarcity”. For next year he does not think capacity will be the problem.
Miettinen’s alternative is edge computing: Apple pulling the models onto devices whose processors are becoming good enough, going to the cloud only when needed. Jokinen answers with NVIDIA: forecasts for GPU compute capacity growth are absurd, but his claim is that they will not suffice — demand will exceed availability specifically on the side of running and training large language models. From there it becomes a question of capitalism: who ends up being able to buy the capacity, small data centre operators or the large cloud providers, when the investments are counted in tens or hundreds of billions.
What the companies offer
Both guests close by describing their businesses. Puupponen does it humbly but directly: Cloud1 keeps a low profile in marketing and sales, but on its narrow patch — Microsoft, data, cloud — it considers itself number one in Finland. “We’re no CGI or Tietoevry, but we probably have the deepest expertise, because this is the only thing we do.”
Jokinen’s pitch is lighter: give Harri some money and you get a dedicated OpenAI API instance. Vuo AI is building tools for its own team to become the world’s best at prompts and at interfaces on top of language models, and he wants to give those to others too: if you want your staff applying and sharing GPT-based tools for sales or HR problem-solving, the email address is [email protected].
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 article is written only from what the transcript contains; the gap has not been filled by inference.
Two of the three speakers are guests, and the MacWhisper transcript carries no speaker labels. Named attributions rest on direct evidence in the transcript — being addressed by name, or a self-reference only one of the guests could make. Passages where the speaker cannot be identified with certainty are marked “one of the guests”.
Watch
The recording lives on the Neuvottelija channel: Tekoäly käyttöön Cloud1 Vuo AI | Harri Puupponen Pasi Jokinen | Neuvottelija 189.