# Taivo Pungas
> Working notes on how technology changes real systems: building with AI, work and judgment, and technology inside organizations.
Public Ghost content for AI and LLM tooling. Use `/llms-full.txt` for consolidated page and post context.
Append `.md` to any post or page URL to get the content in Markdown (for example, `/example-post.md`).

## Pages
- [About](https://www.taivo.ai/about.md) - I'm CIO at Pactum AI, where I support running the company through technology the company depends on. I've seen AI agents from both ends: building them for Fortune 500 teams, and adopting them across a 150-person company. Pactum's AI agents have autonomously negotiated tens of billions of dollars in…
- [Talks](https://www.taivo.ai/talks.md) - Invite me to speak at your event: taivo@pungas.ee. I've been speaking since 2013, mostly at technical and technology-related events in Estonia and around Europe. My typical topics: * my work in building AI and automating human activity at Pactum, Veriff, Starship and elsewhere; * the stories of my…

## Posts
- [Finding talent is a key job](https://www.taivo.ai/finding-talent.md) - I joined Veriff as its first ML engineer, and almost immediately we realised we needed to scale up the AI effort very fast. I hired ten people in four months, and ended up leading the team myself four or five months in. One of those ten was fresh out of university with a master's in physics and alm…
- [A thinking partner is a medium, not a source](https://www.taivo.ai/thinking-partner.md) - A thinking partner's value is in what it pulls out of you, not the text it writes. Five things stock Claude gets wrong, and the 800 words of instructions that fix them.
- [There will always be something impressive left to do](https://www.taivo.ai/there-will-always-be-something-impressive-left-to-do.md) - People worry that if AI keeps improving, there will be nothing impressive left for humans to do. I disagree. What counts as impressive will simply change. A full rewrite of an existing large piece of code today is not trivial, but it's also definitely not impressive anymore. Five years ago, a full…
- [HTTP 402, still reserved](https://www.taivo.ai/http-402-still-reserved.md) - Identity, payments, and accountability online have never really been solved. They're not built into the protocols themselves. Payment was in the web's original vision: Tim Berners-Lee's first HTTP spec already had status code 402, "Payment required", where the client could retry a request "with a s…
- [Great ideas are incubated dumb ones](https://www.taivo.ai/great-ideas-are-incubated-dumb-ones.md) - Estonia announced it will give AI agents digital IDs, and people are kind of skeptical and negative about it. I'm not. I think there's something to this. In general, I try not to shoot down new ideas by arguing against them as a knee-jerk reaction. Many ideas, when born, look quite weak or dumb. Bu…
- [The low-value work left for humans](https://www.taivo.ai/the-low-value-work-left-for-humans.md) - I recently used Claude to do the annual report for my holding company. It's an extremely simple company. There are no meaningful decisions to be made here. One invoice this year, and it's obvious how it should be handled. Still, Claude saved me probably an hour of searching for the correct way of d…
- [Agentic UX ruins the coaching](https://www.taivo.ai/__agentic-ux-ruins-the-coaching.md) - One more thing that's annoying when using AI as a coach: any amount of reasoning tokens will make me switch away. Latency actually matters here. Two seconds is okay; twenty ruins the flow. There is also something more subtle. I don't take as much responsibility for my thoughts, am not as critical.…
- [AI coach](https://www.taivo.ai/__ai-coach.md) - Can you build an AI coach? The obvious answer is yes: AI is great at spotting patterns in what you say, and you can ask it to use any style of coaching, any framework. An exec coach I used to work with thinks you can't build an AI coach, because true change happens in relation to another human. I'v…
- [What to automate, what to keep human](https://www.taivo.ai/what-to-automate-what-to-keep-human.md) - In May 2026 I sat on a panel at Latitude59 in Tallinn called "AI in Your Org: What to Automate, What to Keep Human", with the chief of staff at Hostinger and a co-founder of Sera Leads. Three very different company sizes: 900 people, 150, and 10. Below are my answers, pulled out of the discussion a…
- [How to install a Claude Code skill from GitHub](https://www.taivo.ai/how-to-install-a-claude-code-skill-from-github.md) - Install a Claude Code skill from any GitHub repo: the desktop steps, the two CLI commands, and why Claude keeps asking for permission afterwards.
- [Places to put the AI](https://www.taivo.ai/places-to-put-the-ai.md) - Every product team building with AI faces the same question: where does the AI interaction go? Not the model or the prompt, but the affordance: the surface the user sees and interacts with. For my own reference and discussions I wanted to capture these. Even though the chat box is where this all go…
- [Building a Google Slides renderer with coding agents](https://www.taivo.ai/building-a-google-slides-renderer-with-coding-agents.md) - When I read about the Software Factory concept from StrongDM AI, I was intrigued. In part because of the ability to produce an impressively large and complex component. But even more so with how little it might take: could you really do that simply by taking pre-existing specs from the internet and…
- [Search in program space](https://www.taivo.ai/search-in-program-space.md) - Coding agents run a search in program space. That's the closest analogy I can find for how to use agents productively. If you think of it as an assistant to whom you give tasks, you'll generally be in too tight a loop, giving feedback every few minutes. But the search analogy forces you to consider…
- [Three categories of guardrails](https://www.taivo.ai/three-categories-of-guardrails.md) - There are only three categories of guardrails to prevent harm from agents. First, relying on hard constraints to only allow certain kinds of behaviour. For example, limiting which tokens can be decoded (structured output) or exposing only a specific set of tools to an agent. Assuming correct implem…
- [Week 1: the gap is inside your company](https://www.taivo.ai/week-1-the-gap-is-inside-your-company.md) - I committed to sharing something every day in March about our AI transformation at Pactum. The pace is high, not just internally but in the entire world, so there's a lot that happens and a lot to share. This is the week 1 recap, pulling together what I've learned across the posts, the meetings, an…
- [Core vs edge agents](https://www.taivo.ai/core-vs-edge-agents.md) - As our employees use Claude Code for various tasks, we keep seeing agents repeat the same work. Summarizing a customer's state across Salesforce, email, product analytics, usage data: every agent pulls together the same picture from scratch, every time. Martin Kosk, our enterprise architect, propos…
- [Nobody gets fired for buying Anthropic](https://www.taivo.ai/nobody-gets-fired-for-buying-anthropic.md) - "Nobody gets fired for buying IBM" used to mean: pick the vendor nobody will question. It was a decision driven by fear, not conviction. I think Anthropic is becoming the equivalent for AI, but for a better reason. At Pactum, we didn't choose Anthropic through structured evaluation. We started with…
- [Your AI employees need a handbook](https://www.taivo.ai/your-ai-employees-need-a-handbook.md) - I've seen people try Claude or ChatGPT, get mediocre answers, and conclude AI is overhyped. But the power of agents doing real work for every employee comes from the agent knowing your company, your team, your processes. If you use the vanilla version and expect it to be mind-blowing, you're testin…
- [My agent use cases, part 2](https://www.taivo.ai/my-agent-use-cases-part-2.md) - In part 1 I covered the straightforward agent use cases: copyediting, onboarding, expense reports, Jira management. Those are about doing existing tasks faster. The five below are different. They share a thread: the agent isn't replacing someone's work, it's gathering context I wouldn't have gather…
- [My agent use cases, part 1](https://www.taivo.ai/my-agent-use-cases-part-1.md) - While we're in this mode of rapidly developing agent capabilities, I want to do my part in diffusing knowledge of what agents can do. Since Claude Code IS the software, I am using it more and more for everyday office work, and rapidly discovering new things that agents can (or cannot) do. My daily…
- [Possible](https://www.taivo.ai/possible.md) - People sometimes use the word "impossible" too lightly. When I consider whether something is possible, I consider two angles first: 1. Is it logically provably impossible? 2. Is it against the laws of physics? It's rarely either, at least in my line of work. Putting a problem in these terms makes t…
- [Agent skill - Statistics Estonia](https://www.taivo.ai/__agent-skill-statistics-estonia.md) - Today I published my first agent skill: using Statistics Estonia databases. Often I have some simple question and I know data exists, but I can't be bothered to figure out the clunky UIs and directory trees and subtle variants of tables. This is now automated for me. It's built for Claude Code, but…
- [Claude Code IS the software](https://www.taivo.ai/__claude-code-is-the-software.md) - In 2024, humans created software. In 2025, Claude Code (and Codex, and Cursor, and all the others) created software. In 2026, Claude Code is the software. It's been clear for a while now that the cost of creating software has come down radically, so software becomes disposable: more like a paper to…
- [Vibe coding a music learning app](https://www.taivo.ai/vibe-coding-a-music-learning-app.md) - Reading Shipping at Inference-Speed and with the Christmas holiday on my hands, I again felt a deep desire to ship a side project while getting a feel for state of the art AI coding tools. Usually I temper my enthusiasm for side projects and discard ideas immediately: rarely do I have significant f…
- [Subtractive changes](https://www.taivo.ai/__subtractive-changes.md) - When solving a problem with an existing system, most people tend to add. But often, removing makes a design better. * Dieter Rams's famous principle is "Less, but better." * Strunk & White tell you to "Omit needless words." In fact, the paper People systematically overlook subtractive changes explo…
- [Skills, not agents](https://www.taivo.ai/__skills-not-agents.md) - I used to think shipping an agent product meant building: 1. The core LLM loop. 2. Skills: describing to the LLM how to do particular things (including building API connectors etc). But it seems that Claude Code and Codex are strong enough general-purpose agents that you can just outsource (1) to o…
- [AI Ranch podcast with Evalds Urtans](https://www.taivo.ai/__ai-ranch-podcast-with-evalds-urtans.md) - Mostly talking about AI, Pactum and selling to enterprises. It was recorded at the end of August, so some things are already likely stale! Find the podcast on Youtube or your favourite podcast app. Excerpts: I’m not very surprised or disappointed by GPT-5, and that’s because the trajectory is actua…
- [Principal](https://www.taivo.ai/__principal.md) - I recently noticed that it's a bit clunky to talk about "users" of our agents. Building complex enterprise software, we have multiple types of users, from operations teams making daily decisions to executives looking at high level dashboards, and of course suppliers. But often none of these people…
- [Anchor on what won't change](https://www.taivo.ai/__anchor-on-what-wont-change.md) - When things are stable, you make plans by anchoring to the question "what will change?". With your customers, your company, technology, regulation or anything else about the world. This does not work when things are changing fast. If best-in-class tech changes every 6 months, or if your product str…
- [Invest into context](https://www.taivo.ai/__invest-into-context.md) - I have a 14,000-word document I maintain manually, to add into ChatGPT queries I make related to my role as CTO. It covers a lot of ground about Pactum: * An extremely basic description of who we are and what we do * Company strategy, long-term goals, financial model, and progress towards those * O…
- [My deep research use cases](https://www.taivo.ai/__my-deep-research-use-cases.md) - Deep research is underappreciated. The feature exists in all three LLM chat apps I pay for (ChatGPT, Claude and Gemini) and is unfortunately heavily rate limited, which makes sense given the likely token cost. But the massive amount of tokens burnt on reading and analyzing hundreds of articles is t…
- [The complements to AI will increase in value](https://www.taivo.ai/__the-complements-to-ai-will-increase-in-value.md) - Here's a model I use to think about how AI will change work. Some parts of every activity will cost 1,000x less than they do when a human does it. For engineers it might mean reading and typing code, for radiologists reading X-ray images, etc. But it's not 100% of the time a human spends on that jo…
- [Shaping is high-leverage work for leaders](https://www.taivo.ai/__shaping-is-high-leverage-work-for-leaders.md) - I've recently seen several engineers struggle when moving to the role of leading a small product team and managing a couple engineers. There are so many newly important meetings and Slack messages (that you could ignore as an individual contributor) and they take up most of your time, but don't fee…
- [Are software jobs set to grow or shrink?](https://www.taivo.ai/__are-software-jobs-set-to-grow-or-shrink.md) - There seem to be two views of what will happen with engineers' jobs given the new efficiencies from using advanced AI coding tools (Cursor, Claude Code, etc). The first view is partial rebound. Fewer engineers will be needed in total: if engineers are 20% more effective per hour, there will be some…
- [Gemini 2.5 Pro isn't quite Claude 3.7 at coding](https://www.taivo.ai/__gemini-2-5-pro-isnt-quite-claude-3-7-at-coding.md) - Gemini 2.5 Pro was just released and it could be a big deal, if its coding abilities pan out. The current positioning of Gemini has been roughly that it is a tad behind the OpenAI/Anthropic models of the same class, and far behind the coding capabilities of Claude 3.7 Sonnet which is considered unm…
- [Hiring as the marriage problem](https://www.taivo.ai/__hiring-as-the-marriage-problem.md) - At one point in my career, our company was hiring for role of a Head of Marketing. I got a strong referral from a friend, reached out, the match was great, and she passed the interview process with ease. And then we just... sat on it. The hiring manager said he wanted to consider three candidates a…
- [Which European institution is running a massive LLM hackathon?](https://www.taivo.ai/__which-european-institution-is-running-a-massive-llm-hackathon.md) - Anthropic, OpenAI and possibly others are doing one in the US with 1,000 participants from public sector research institutes. It's called an "AI Jam" and the companies are obviously doing it out of commercial interest. Even so, the government sector must have such obvious use cases. Just running a…
- [Will AI take us towards refinement of the self?](https://www.taivo.ai/__will-ai-take-us-towards-refinement-of-the-self.md) - In Would You Rather Have Married Young?, Lillian Fishman writes about the "fundamental ethos that had long governed the young secular woman" in the last 50 years: Experience, we hoped, would broaden us. The new object seems to be the inverse: the contraction and refinement of the self, within and a…
- [Half AI, half random curiosities](https://www.taivo.ai/__half-ai-half-random-curiosities.md) - I rarely look at the search analytics for my blog, but today I did. I am really amused by the most common search queries that bring people to this website, which definitely show the variety of content I post here. Here are the top ones in the past month, with similar ones grouped: 1. are LLMs deter…
- [Vibe-code with stable infrastructure](https://www.taivo.ai/__vibe-code-with-stable-infrastructure.md) - Vibe-coding produces close-to-unmaintainable code right now. But it is an acceptable trade-off in places where it produces decent results and maintainability matters less -- for example making simple UIs. So maybe a good approach is to vibe-code with stable infrastructure. This riffs off of Faceboo…
- [Negotiating against an LLM](https://www.taivo.ai/negotiating-against-an-llm.md) - Writing this post I got heavy AI assistance, starting off with my raw notes of the competition. I know it's not up to my usual quality – this was the only way I could find time to share this. Overall I consider the result not good enough, but I will keep experimenting with different kinds of AI ass…
- [The GPU export limit might start to matter](https://www.taivo.ai/__the-gpu-export-limit-might-start-to-matter.md) - White House's new chip rules seem to not have huge impact right now, in the "tier 2" countries like Estonia. It is unlikely we ever would have been able to build a $1B data center anyway. But this restriction could still become significant. I don't fully understand but I assume the per-country limi…
- [Good LLM devtools are also good human devtools](https://www.taivo.ai/__good-llm-devtools-are-also-good-human-devtools.md) - What does a well designed developer tool (programming language, library, API, CLI utility) look like -- for an LLM? The more we write code with Copilot and Cursor and chat-assistants, the more important this question is. Off the top of my head, a good tool for LLM programmers: * Chunks, abstracts a…
- [LLMs highlight lack of progress in real world](https://www.taivo.ai/__llms-highlight-lack-of-progress-in-real-world.md) - With LLMs now multimodal and extremely cheap, they can do much of what a human can, when limited to only the virtual. This rapid progress in the world of bits, however, is a sore reminder of how slowly things improve in the physical world. I need an AI that can: * plan meals, manage food inventory,…
- [Creating single-file apps with LLMs](https://www.taivo.ai/__creating-single-file-apps-with-llms.md) - To prototype a mobile-optimized writing app I used LLMs as code generators. The idea isn't new but I got inspiration from reading Simon Willison's experiments creating micro-UIs. The idea is to prompt for whatever functionality you need, and ask for a single HTML file containing CSS and JS as neede…
- [Writing on mobile is different](https://www.taivo.ai/__writing-on-mobile-is-different.md) - Whenever I write here, I do it on my laptop, almost never on the phone. I do have a Bluetooth keyboard that connects to my phone, but it's rare that I remember to bring it with me, and have a moment to take it out of the bag. Plus it's un-ergonomic. But I also almost never have deep focus time for…
- [DIY parametric wall art: plans to production](https://www.taivo.ai/parametric-wall-art.md) - Watching Youtube woodworking videos is my guilty pleasure. But until now, I'd had little opportunity to make something myself: no space for a workshop in my small apartment and a busy family/work schedule haven't really allowed it. Recently on summer vacation I had an idea though: I could outsource…
- [Best pastry in Tallinn](https://www.taivo.ai/__best-pastry-in-tallinn.md) - This is my running ranking of the best buns in Tallinn, updated as bakeries open and close. The latest shake-up came in March 2026, when long-time leader Sumi closed its doors. 1. Pulla Bakery Location: Old Town, Voorimehe 7. Price: 3€ per bun. The front of this tiny cafe is just a few windows alon…
- [Frontier LLMs come every 1.5-2 years](https://www.taivo.ai/__frontier-llms-come-every-1-5-2-years.md) - I have a theory that the LLM frontier moves every 1.5-2 years. Let me qualify. I mean that a major leap of the whatever is the best model (currently GPT-4) happens with that interval. Incremental improvements don't count: e.g. Claude 3 Opus is claimed to be maybe slightly better than GPT-4, but it…
- [Explaining quiet quitting](https://www.taivo.ai/__explaining-quiet-quitting.md) - Quiet quitting is a recent name for something I am sure has been happening for all of time. It probably has something to do with the person's life circumstances, or something about the macro environment being depressing, or whatever. But let me propose a simple economic justification. Consider the…
- [SAD lights](https://www.taivo.ai/__sad-lights.md) - It seems like everyone I talk to recently is thinking in the same direction: managing their seasonal affective disorder, or SAD. If you live in Estonia, or anywhere with a long winter, or maybe even the less sunny parts of California, you'll know what I'm talking about. If you don't, the TL;DR is t…
- [Strive for off-grid discipline](https://www.taivo.ai/__strive-for-off-grid-discipline.md) - We lost a floorball game yesterday. I can't stop thinking about it. Maybe because it felt like a pointless loss, like we lost because of something fully within our grasp. What happened? Two periods into the game, we were ahead 3:1. We had kept our defence intact, our game decent. But then in the th…
- [Momentum](https://www.taivo.ai/__momentum.md) - An object that has a lot of momentum is hard to stop. A bowling ball. An ocean liner. A person who will not allow themselves to be derailed. Physically, an object of any amount of momentum could be stopped in a very short time, if enormous force is exerted. And the same could apply, in theory, to m…
- [Non-judgemental awareness is curative](https://www.taivo.ai/__non-judgemental-awareness-is-curative.md) - Rather than try to fix things, all you have to do is notice, non-judgementally, that you’re doing them. Even while you’re involved in this non-judgemental noticing, you will notice a barrage of impulses to try different things, to intervene, to try, to fix. That’s fine, just notice them non-judgeme…
- [Curiosity arises from lack of feed](https://www.taivo.ai/__curiosity-arises-from-lack-of-feed.md) - Here's a hypothesis: I think I'll always find something I am curious about. I am a naturally curious person, and I would actually guess that everyone is? It seems impossible that there would ever be a moment where nothing could interest me. But I think curiosity arises from lack of Dopamine. Basica…
- [Are LLMs deterministic?](https://www.taivo.ai/__are-llms-deterministic.md) - No. You can see for yourself: setting the temperature variable to 0 (meaning you always sample the most likely token from the output distribution), you'd expect GPT-3.5 and 4 to produce the same output every time you call them. However, they don't. Why do LLMs produce different outputs across diffe…
- [Recent LLM launches, and LLM rumors](https://www.taivo.ai/__recent-llm-launches-and-llm-rumors.md) - Llama 3 is already training according to Zuck. There are conflicting sources & rumors, and the release date claims vary across all of 2024. For GPT-5 there are even no reliable rumors; if training started within the past few months then my back of napkin is that it may be launched early 2025, becau…
- [A wild speed-up from OpenAI Dev Day](https://www.taivo.ai/__a-wild-speed-up-from-openai-dev-day.md) - I'll share more thoughts on OpenAI Dev Day announcements soon, but one huge problem for any developer is LLM API latency. And boy, did OpenAI deliver. On a quick benchmark I ran: * gpt-4-1106-preview ("gpt-4-turbo") runs in 18ms/token * gpt-3.5-turbo-1106 ("the newest version of gpt-3.5") runs in j…
- [LLM+API vs LMM+UI](https://www.taivo.ai/__llm-api-vs-lmm-ui.md) - The two most famous startups focused on making Agent middlewares seem to be Imbue and Adept. Both companies' goal is to have a large model use a computer effectively, but it is interesting how they seem to bet on two different approaches. * Imbue: use LLMs (text-based) with API-based tools. * Adept…
- [Simplicity is essential in a generative world](https://www.taivo.ai/__simplicity-is-essential-in-a-generative-world.md) - * "less is more" (proverb) * "perfection is finally attained not when there is no longer anything to add, but when there is no longer anything to take away" (Antoine de Saint Exupéry) * "omit needless words" (Strunk & White) * "It is vain to do with more what can be done with fewer" (Occam's razor)…
- [Diverge-converge cycles in LLMs](https://www.taivo.ai/__diverge-converge-cycles-in-llms.md) - The Double Diamond is, roughly, a design framework consisting of 4 steps: 1. Diverging on problem (Discover). Explore widely to gather a broad range of insights and challenges related to the problem. 2. Converging on problem (Define). Analyze and synthesize the gathered insights to define a clear a…
- [LLM latency is linear in output token count](https://www.taivo.ai/__llm-latency-is-linear-in-output-token-count.md) - All top LLMs, including all GPT-family and Llama-family models, generate predictions one token at a time. It's inherent to the architecture, and applies to models running behind an API as well as local or self-deployed models. Armed with this knowledge, we can make a very accurate model of what the…
- [RAG is more than just embedding](https://www.taivo.ai/__rag-is-more-than-just-embedding.md) - 90% of time when people say "Retrieval-augmented generation" they mean that the index is built using an embedding model like OpenAI's text-embedding-002 and a vector database like Chroma, but it doesn't have to be this way. Retrieval is a long-standing problem in computer science -- a couple of PhD…
- [Retrieval-augmented generation](https://www.taivo.ai/__retrieval-augmented-generation.md) - Retrieval-augmented generation, or RAG, is a fancy term hiding a simple idea: Problem: LLMs can reason, but they don't have the most relevant facts about your situation. They don't know the location of your user, or the most relevant passage from the knowledge base, or what the current list of cust…
- [Properties of a good memory](https://www.taivo.ai/__properties-of-a-good-memory.md) - Apps with no Memory are boring. Compare a static website from the 90s with any SaaS or social network or phone app: the former knows nothing about you, the latter knows a lot. From UI preferences (dark or light mode?) to basic personal data (name?) to your friend list to what things you engage with…
- [Things I've underestimated - Sep 2023](https://www.taivo.ai/__things-ive-underestimated-sep-2023.md) - After attending the Ray Summit in San Francisco this week, I realized I had previously discounted several interesting things. Here's what I now want to explore more. Semantic Kernel I've gotten so used to langchain that I haven't really considered switching... all the while loving to hate it. As I…
- [Launching OpenCopilot](https://www.taivo.ai/__launching-opencopilot.md) - Across the many experiments I've made this year (and which I've written about here) I've felt the need for better tools. Specifically, for the past few months I have been building copilots, and doing so from scratch takes a bunch of work every time. So we decided to release OpenCopilot: an OSS fram…
- [Context engineering is information retrieval](https://www.taivo.ai/__context-engineering-is-information-retrieval.md) - The stages of an LLM app seem to go like this: * Hardcode the first prompt, get the end-to-end app working. * Realise that the answers are bad. * Do some prompt engineering. * Realise the answers are still bad. * Do some more prompt engineering. * Discover vector databases!!!1 * Dump a ton of data…
- [Making GPT API responses faster](https://www.taivo.ai/__making-gpt-api-responses-faster.md) - GPT APIs are slow. Just in the past week, the OpenAI community has had 20+ questions around that. And not only is it rare for users to tolerate 30-second response times in any app, it is also extremely annoying to develop when even basic tests take several minutes to run. (Before diving into these…
- [agentreader - simple web browsing for your Langchain agent](https://www.taivo.ai/__agentreader-simple-web-browsing-for-your-langchain-agent.md) - This is a short link-post to a new repo I just released. While working on Why AutoGPT fails and how to fix it I created a handy web-browsing Tool for langchain Agent middlewares and now finally got around to open-sourcing it. Here is the repository: github.com/taivop/agentreader.
- [Why AutoGPT fails and how to fix it](https://www.taivo.ai/__why-autogpt-fails-and-how-to-fix-it.md) - A couple weeks after AutoGPT came out we tried to make it actually usable. If you don't know yet, it looks amazing on first glance, but then completely fails because it creates elaborate plans that are completely unnecessary. Even asking it to do something simple like "find the turning circle of a…
- [Core innovations of AutoGPT](https://www.taivo.ai/__core-innovations-of-autogpt.md) - AutoGPT (repo) went viral on Github and looks impressive on Twitter, but almost never works. In the process of trying to improve it I dug into how it works. Really there are two important parts to AutoGPT: a plan-and-execute workflow, and looped Chain-of-thought (CoT). Plan-and-execute Say the user…
- [GPT-3.5 and GPT-4 response times](https://www.taivo.ai/__gpt-3-5-and-gpt-4-response-times.md) - Some of the LLM apps we've been experimenting with have been extremely slow, so we asked ourselves: what do GPT APIs' response times depend on? It turns out that response time mostly depends on the number of output tokens generated by the model. Why? Because LLM latency is linear in output token co…
- [How data is used for LLM programming](https://www.taivo.ai/__how-data-is-used-for-llm-programming.md) - Software 1.0 -- the non-AI, non-ML sort -- extensively uses testing to validate things work. These tests are basically hand-written rules and assertion. For example, a regular expression can be easily tested with strings that should and should not give a match. In software 2.0, and specifically sup…
- [Hacky multimodality](https://www.taivo.ai/__hacky-multimodality.md) - GPT-4 supports images as an optional input, according to OpenAI's press release. As far as I can tell, only one company has access. Which makes you wonder: how can you get multimodality support already today? There are basically two ways for adding image support to an LLM: 1. Train a vision encoder…
- [First thoughts on AI moratorium](https://www.taivo.ai/__first-thoughts-on-ai-moratorium.md) - Context: first thoughts on Pause Giant AI experiments. I will refine my thinking over time. * I had not thought about AI safety much since ~2017, after thinking a lot about it in 2014-2017. In 2017, I defended my MSc thesis on an AI-safety-inspired topic (though very narrow and technical in nature)…
- [Agents are self-altering algorithms](https://www.taivo.ai/__agents-are-self-altering-algorithms.md) - Chain-of-thought reasoning is surprisingly powerful when combined with tools. It feels like a natural programming pattern of LLMs: thinking by writing. And it's easy to see the analogy to humans: verbalizing your thoughts in written (journalling) or spoken ("talking things through") form is a good…
- [Index to reduce context limitations](https://www.taivo.ai/__index-to-reduce-context-limitations.md) - There is a very simple, standardized way of solving the problem of too small GPT context windows. This is what to do when the context window gets full: Indexing: 1. Chunk up your context (book text, documents, messages, whatever). 2. Put each chunk through the text-embedding-ada-002 embedding model…
- [Langchain frustration](https://www.taivo.ai/__langchain-frustration.md) - I was building an agent with langchain today, and was very frustrated with the developer experience. It felt hard to use, hard to get things right, and overall I was confused a lot while developing (for no good reason). Why is it so hard to use? Here are some of my observations. 1. Verbose. If I ha…
- [Office 365 copilot is extremely impressive](https://www.taivo.ai/__office-365-copilot-is-extremely-impressive.md) - Microsoft recently held an event where they announced the "Office 365 Copilot". It was extremely impressive to me, to the extent that (when this launches) I will consider switching from Google's work suite to Microsoft's. Why is this announcement so impressive? In one word, integration. Pretty much…
- [Your non-AI moat](https://www.taivo.ai/__your-non-ai-moat.md) - "What's the Moat of your AI company?" That seems to be top of mind for founders pitching their novel idea to VCs -- and the most common question they get. But I think it might not be the right question. Right now, nobody seems to have a good answer. a16z doesn't. At the same time, the pace of new b…
- [GPT-4, race, and applications](https://www.taivo.ai/__gpt-4-race-and-applications.md) - GPT-4 came out yesterday and overshadowed announcements, each of which would have been bombshell news otherwise: * Anthropic AI announcing their ChatGPT-like API -- likely the strongest competitor to OpenAI today (only waitlist) * Google announcing PaLM API (currently it's just marketing -- no publ…
- [Alpaca is not as cheap as you think](https://www.taivo.ai/__alpaca-is-not-as-cheap-as-you-think.md) - "Alpaca is just $100 and competitive with InstructGPT" -- takes like this are going around Twitter, adding to the (generally justified!) hype around AI models. It is indeed a very encouraging result. Specifically, that it took so little compute to train something that achieves competitive results o…
- [Chain-of-thought reasoning](https://www.taivo.ai/__chain-of-thought-reasoning.md) - Matt Rickard has a concise overview of Chain-of-thought: the design pattern of having an LLM think step by step. To summarize, the four mentioned approaches from simpler to more nuanced are: 1. Add "Let's think step-by-step" to the prompt. 2. Produce multiple solutions, have the LLM self-check on e…
- [Define a vector space with words](https://www.taivo.ai/__define-a-vector-space-with-words.md) - I've been experimenting with embedding ideas into GPT-space, and using the resulting vectors to visualize. For example, you could plot different activities based on two axes: how safe they are, and how adrenaline-inducing they are. How does it work? The intuition is the following. In words, you can…
- [AI product overhang](https://www.taivo.ai/__ai-product-overhang.md) - There is a massive AI overhang today. The term is analogous to snow overhang. When snow slides down a roof, sometimes a section goes over the edge and is seemingly supported by nothing. You know it will eventually fall; the tension is in the air. It's a question of time. I have the same feeling wit…
- [Self-driving apps](https://www.taivo.ai/__self-driving-apps.md) - There's a now-famous classification of different levels of self-driving: * Level 0: No automation * Level 1: Driver assistance * Level 2: Partial automation * Level 3: Conditional automation (human fallback always available) * Level 4: High automation (can safely pull over) * Level 5: Full automati…
- [Text is the interface, not the use case](https://www.taivo.ai/__text-is-the-interface-not-the-use-case.md) - Probably 80% of (mainstream, non-ML) Twitter is excited about the potential of GPT and ChatGPT. The other 20% is skeptical. Few are indifferent. But both the optimists and pessimists seem to focus on the obvious set of use cases: text generation and chatbots. That is a failure of imagination. Text…
- [Prototyping with GPT API is fun](https://www.taivo.ai/__prototyping-with-gpt-api-is-fun.md) - I've built around 10 prototypes with GPT over the past couple weeks and it's surprisingly fun. Why is that? Obviously it's a shiny new toy. Whenever I discover a new tech capability, I get excited. I recall discovering D3.js, Bokeh, Keras... These appeal to the Technology enthusiast in me. But ther…
- [How do you decompose knowledge work for AI](https://www.taivo.ai/__how-do-you-decompose-knowledge-work-for-ai.md) - I hacked a bit on an AI experiment: idea generator. In general, working on these AI apps makes you think a lot about the work process and unique value add of Knowledge workers in general, and myself specifically. How predictable is it? Which parts are easy to automate; which are easy to assist? How…
- [AI shovels don't beat AI apps](https://www.taivo.ai/__ai-shovels-dont-beat-ai-apps.md) - NFTPort sells tools in web3. I've heard people say that this is a smart approach in a volatile market, because selling shovels in a gold rush is safer than looking for gold yourself. That seems wrong. Sure, during the rush your revenue is more stable, but you still depend on that market's macro cyc…
- [Hands-on integrating NFTPort with Unity](https://www.taivo.ai/hands-on-integrating-nftport-with-unity.md) - This is a talk I gave in Lisbon in October 2022, at IPFS camp. What makes it somewhat special: I generated all the illustrations for this deck using DALL-E as an experiment. It actually is a decent way of making pretty slides; I think with a better UI this would be a major improvement in my slide-m…
- [ETH New York workshop: Launching your NFT app with NFTPort](https://www.taivo.ai/eth-new-york-workshop-launching-your-nft-app.md) - This is a short NFTPort workshop I recorded for the ETHGlobal NY hackathon. I was actually in New York around that time but sadly had to fly out before the in-person event itself took place. Recommended over the last one because this is pre-recorded and heavily edited, so much more information-dens…
- [Announcing NFTPort's $26M Series A](https://www.taivo.ai/the-potential-of-nfts-announcing-nftports-26m-series-a.md) - Today, we announce a $26M fundraise with NFTPort. You may be surprised. This wouldn't make much sense if NFTs were only about buying and selling images of apes. Art and profile pictures may fetch high prices when sold as NFTs, but it feels like those things lack some tangible "real world" value tha…
- [GPT and Me](https://www.taivo.ai/gpt-and-me.md) - DALL-E 2 and Imagen are the hot new models everyone is sharing: the eye candy is undeniable. But since the original release of the GPT-3 large language model, OpenAI has quietly released the GPT API and playground for anyone to try. Here's a collection of my experiments with the most recent version…
- [20.1% yearly inflation in Estonia](https://www.taivo.ai/estonia-inflation-2022.md) - What did Estonians start paying more for in April 2022?
- [HBS vs BAYC](https://www.taivo.ai/hbs-vs-bayc.md) - At a New Year's Eve house party a few years ago, I met someone who had graduated from Harvard Business School. Among other topics, I brought up something I couldn't figure out: why would someone pay $200,000 [https://www.hbs.edu/mba/Pages/cost-of-attendance-class-of-2023.aspx] for something you cou…
- [Thoughts from my first web3 conference](https://www.taivo.ai/thoughts-from-devconnect-amsterdam.md) - In the spirit of documenting my thought process instead of already-formed thoughts, here are my impressions from the recent DevConnect – a major week-long gathering of the web3 developer community in Amsterdam. * Many web3 companies are default-global. Most don't mention where they are based, becau…
- [Workshop: building NFT apps with NFTPort](https://www.taivo.ai/building-nft-apps-with-nftport.md) - This is a short workshop I gave at the Road to Web3 online hackathon: a brief demo of NFTPort [https://nftport.xyz/]'s APIs to people who have a rough understanding of what NFT [https://www.taivo.ai/notes/__nft/]s are. If you are not one of those people, the original NFT standard spec (ERC-721) [ht…
- [Startup stock options: a beginner's guide](https://www.taivo.ai/startup-stock-options.md) - Rationally valuing equity for startup employees
- [Web3: what I'm doing next](https://www.taivo.ai/life-update-nftport.md) - As I wrote before, I ended my startup journey [https://www.taivo.ai/startup-year] in November and started looking for a job. A lot has happened between then and now. I've spoken to tons of people, looked at a broad range of options, and decided to join one early-stage company. It felt weird to anno…
- [Foundering founder: the story of my startup year](https://www.taivo.ai/startup-year.md) - How I spent a year trying to start a startup -- and failed.
- [Product before problem](https://www.taivo.ai/product-before-problem.md) - You might imagine all great startups started from a clear mission. Sometimes they were, but usually, that story is woven in hindsight to cover up the gnarly pivots of the early days. In PR-optimised founding stories, the narrative usually goes like this. Our visionary founders found an apparent $PR…
- [Diffuse mode: thinking by relaxing](https://www.taivo.ai/diffuse-mode.md) - I recently started drinking coffee again. Caffeine has always had the effect of narrowing my focus and reducing mind-wandering. The downside is, caffeine makes me less likely to get into diffuse mode, the other main mode of thinking. Focused mode means taking a direct, head-on approach to work: wri…
- [Tesla's data engine: the road to full self-driving](https://www.taivo.ai/tesla-data-engine.md) - I've driven a Tesla only once but looking at Karpathy's recent presentation at the CVPR conference, soon no human will. I think Tesla's unique data engine is what will get them there. In their self-driving stack, Tesla has always been betting on cameras and radar. This contrasts with Waymo's bet on…
- [Becoming a definite optimist](https://www.taivo.ai/definite-optimist.md) - Peter Thiel has [https://www.goodreads.com/book/show/18050143-zero-to-one] a 2x2 matrix: indefinite vs. definite, and optimist vs. pessimist. I want to be a definite optimist: imagine a better future and work to build it. Recently, I've been more of an indefinite optimist. * Hoping that startup ins…
- [If you can measure it, you can improve it](https://www.taivo.ai/measure-improve.md) - Early this year, I noticed I'd put on some lockdown-weight. I wasn't super worried but wanted to get back to my typical, healthier weight level. My first intuition in this situation was, "let me go on a diet." Diets can work. But they are expensive: you spend willpower every day on avoiding the del…
- [Video: Invest in Estonia, 1993](https://www.taivo.ai/invest-in-estonia-1993.md) - This is a hidden gem: Estonia's pitch as a startup country, almost 30 years ago. As far as I know, it hasn't been on the Internet yet. This video (and an accompanying paper publication) was distributed in 1993 by the Estonian Privatisation Agency to invite foreign investors to Estonia. A few highli…
- [Integrate work and life](https://www.taivo.ai/integrate-work-and-life.md) - You could separate work from life. As in the phrase "work-life balance", work versus life. They can be separate buckets with separate goals, time slots, people, and emotions. Separation definitely makes sense. Work lets us pay for rent and groceries, but might not be fulfilling. Time with friends,…
- [The negotiation technique that enables deep conversations](https://www.taivo.ai/negotiation-technique-conversations.md) - The single biggest thing that made me a better friend, partner, colleague, and mentor came from an unexpected place. Two years ago, I was working through a breakup and rethinking how I approach all relationships. At a conference in Berlin, I happened to participate in a Circling workshop. It's is e…
- [Stock options are hard](https://www.taivo.ai/stock-options-are-hard.md) - As an employee, startup stock options are hard. I'm feeling pretty confident by now, but only because I've seen my friends get burned and been burned myself. What's so hard? 1. At one company, my option contract was to be signed "at a future date", not together with my employment contract. In the e…
- [Commitment](https://www.taivo.ai/commitment.md) - I'm both a millennial and a startup founder and both of these groups are often considered flaky perhaps, unable to commit. They seem like they're, flip-flopping between startup ideas or between jobs, between relationships. And I think all of this comes from different expectations to commitment. So…
- [Cause variance](https://www.taivo.ai/cause-variance.md) - How do you end up with an outlier success? I guess the common advice would be to "leave your comfort zone". That's hard to do. In the comfort zone, everything makes sense. Things work. Clients are using the product. You might even be growing slightly. It would be stupid to mess with something that…
- [Digital dopamine-seeking](https://www.taivo.ai/digital-dopamine-seeking.md) - My nervous system is in a civil war over my actions. The dopamine-seeking side keeps winning. Dopamine itself is crucial: it's a neurotransmitter driving you to achieve external goals. The bad part is being driven to the easiest, least meaningful ones. I try not to feed the enemy. Soft-blocking add…
- [School-career mashup](https://www.taivo.ai/school-career-mashup.md) - There's an abrupt life transition almost everyone goes through: school to job. You might go from middle school to manual labour, or PhD to teacher. But regardless of details, a more gradual shift would be better. Most entry-level jobs could be apprenticeships that mix education and work. Why mix th…
- [Courage in the face of variance](https://www.taivo.ai/courage-in-the-face-of-variance.md) - Startups have high variance: the top few % reap most of the rewards. The same is true for artists of all kind, and to a lesser extent, software engineers: some Waymo engineers earn 100x the salary of the lowest-paid programmers. Risk is usually taken to be bad. But it might not be if you can influe…
- [The office](https://www.taivo.ai/the-office.md) - Who needs an office these days? It’s a needless expense. I currently work alone, and when I meet others I do it via video call or on a walk outside. Plus, private offices are expensive. A desk in a co-working space is 150€ — and not an improvement over my working-from-home situation — and private r…
- [To kill a mock-up, first](https://www.taivo.ai/to-kill-a-mock-up-first.md) - I ran my first design sprint last week. According [https://www.gv.com/sprint/] to the creators: > The sprint is a five-day process for answering critical business questions through design, prototyping, and testing ideas with customers. The starting point of a sprint could be a problem like “website…
- [A productive creative](https://www.taivo.ai/a-productive-creative.md) - I recently reflected on my productivity and discovered, to my horror, that I’m doing pretty badly! I used to be very organized and do lots of creative, important work consistently every day for months on end. Somehow, I had lost that in the past 3-5 years. Doing what comes naturally is a lousy meth…
- [More pre-trained models, please](https://www.taivo.ai/more-pre-trained-models-please.md) - Every classification task is linearly separable in the right feature space. This statement is a clue to building ML solutions that scale to many different use cases. Here’s a visual explanation. Say you have data points in a given two-dimensional vector space. If you can use a straight line to sepa…
- [It's rocket science](https://www.taivo.ai/rocket-science.md) - “Founders at Work [https://www.goodreads.com/book/show/98233.Founders_at_Work]” is a unique book by Jessica Livingston, one of the founders of YCombinator. It contains un-narrated, almost unedited interviews with founders of successful startups like Paypal, Hotmail, Apple, and 27 others. I’m only t…
- [Datasets: the source code of Software 2.0](https://www.taivo.ai/datasets-source-code.md)
- [Datasets carve the terrain of AI](https://www.taivo.ai/dataset-terrain.md) - Lately, Twitter has been full of the 2020 US election, which has displaced everything interesting. That means it’s a good time to blog. I’ve recently been reading James C. Scott’s Seeing Like a State, which manages to combine forestry, agriculture, and land/city planning into a history of top-down…
- [Creating AI is Curating Examples](https://www.taivo.ai/creating-ai-is-curating-examples.md) - A few years ago at Starship, I contributed to the Data is the Specification [http://dataisspec.github.io/] manifesto. The core idea is that it's better to solve problems directly against a collection of examples, as opposed to trying to generalise the problem first and then solving the general prob…
- [How do AI companies earn money?](https://www.taivo.ai/how-do-ai-companies-earn-money.md) - Which business models benefit from AI? This is an audience question I recently got at a webinar aimed at early-stage startup founders. The answer is almost like enumerating all business models because AI is like Javascript. Which businesses could benefit from Javascript? It can create value at almo…
- [Two meanings of "AI"](https://www.taivo.ai/ai-meaning.md) - What do people mean by “artificial intelligence”? When I say “AI” I mean one of two things. First it could be the experience of a product feeling intelligent. Or I could refer to what it looks like technically: that the system uses machine learning, deep neural networks, code where you optimise par…
- [Your AI team needs DataOps](https://www.taivo.ai/dataops.md) - A startup’s AI work often starts from a developer hacking algorithms on the side. When a generalist engineer with no data science background works on prediction problems, they often don’t use a dataset at all, or at best a small one. For example, in the early days of Veriff, a front-end developer t…
- [How to build your AI startup](https://www.taivo.ai/how-to-build-your-ai-startup.md)
- [Timid New World](https://www.taivo.ai/timid-new-world.md) - As the coronavirus pandemic escalated, I was travelling around the world. The timing was unfortunate, but the endless hours of flying also presented an opportunity for reflecting on the virus. Most people correctly focus on the immediate effects of the virus: survival is required for anything inter…
- [ML at Veriff: What is possible in a year](https://www.taivo.ai/machine-learning-at-veriff-what-is-possible-in-a-year.md) - devclub.lv meetup 07.11.2019ML at Veriff: What is possible in a year?Google Docs

## Optional
- [RSS Feed](https://www.taivo.ai/rss/)
- [Sitemap](https://www.taivo.ai/sitemap.xml)
- [Full content of pages and posts](https://www.taivo.ai/llms-full.txt)