Sign in

Tommi Somersuo

@tommis.fi
2.2K followers 827 following 1.4K posts

AI Tech Lead @ Solita.fi | Agentic engineering, architecture & governance. Offline: Runner, angler, RES/SRA, Tibetan Terrier owner.

PostsRepliesMedia
Tommi Somersuo @tommis.fi · 18/09/2026
prismml.com/news/bonsai-...: Introducing Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint
000
Tommi Somersuo @tommis.fi · 13/08/2026
Overview of SkillZip. One-shot compression first recovers the skill contract, then applies the “explain once, reference many” principle to repeated rules and workflows, while unique and uncertain content remains explicit. Zip-on-Write compares each patch with the affected contract neighborhood and performs occasional repacking when reuse accumulates across patches.
000
Tommi Somersuo @tommis.fi · 04/08/2026
Managing Docker telemetry 2030:
010
Tommi Somersuo @tommis.fi · 24/07/2026
Fetch https://math.mino.mobi/conjectures (table is JS-rendered; crawl
the linked context pages for rows). For every entry tagged plain
"counterexample" — skip "deep" and "existence" — spawn an isolated
agent that first writes an executable exact-arithmetic verifier for a
candidate refuting object (if it can't, it stops), then authors its
own breakthrough prompt naming a concrete structural ansatz (in the
style of "Construct a counterexample to the general non-planar case
of X. You should do a breakthrough and find a structured
counterexample."), then loops propose → verify → log failed ansatz to
a ledger, never repeating a ledger entry, until the verifier passes
or a disk-based iteration cap the agent cannot edit runs out. Success
is only ever declared by the verifier. Report every terminal state;
nothing merges without my review.
020
Tommi Somersuo @tommis.fi · 21/07/2026
180
Tommi Somersuo @tommis.fi · 11/07/2026
+
070
Tommi Somersuo @tommis.fi · 03/06/2026
microsoft.ai/news/introdu...
000
Tommi Somersuo @tommis.fi · 03/06/2026
microsoft.ai/news/introdu...
100
Tommi Somersuo @tommis.fi · 03/06/2026
New Microsoft AI models: microsoft.ai/news/buildin...
110
Tommi Somersuo @tommis.fi · 22/04/2026
Visual Studio Code Agents is a new preview companion app that ships alongside VS Code Insiders: code.visualstudio.com/updates/v1_1...
010
Tommi Somersuo @tommis.fi · 13/04/2026
x.com/marmaduke091...
000
Tommi Somersuo @tommis.fi · 03/04/2026
Claude Code 2.1.9.1 includes likely fix for token consumption bug: github.com/anthropics/c...
110
Tommi Somersuo @tommis.fi · 02/04/2026
docs.z.ai/guides/vlm/g...
051
Tommi Somersuo @tommis.fi · 24/03/2026
Typoglycemia is a thing:
A plain three-column, nine-row table on a white background with no headers or borders, containing 27 typoglycemic variations of the word “Dependabot”
120
Tommi Somersuo @tommis.fi · 04/03/2026
claude.com/blog/improvi...
010
Tommi Somersuo @tommis.fi · 31/01/2026
hypercontext.sh : Hypercontext is a knowledge skill for Claude Code. Invoke /hypercontext and the agent introspects — threads, files, tools, remaining context — and renders a spatial ASCII map.
132
Tommi Somersuo @tommis.fi · 19/01/2026
This implies that Hobbits were the OG 10x engineers.
Hobbit meal times schedule
020
Tommi Somersuo @tommis.fi · 19/01/2026
Why does Edge ask this even with all these settings disabled?
010
Tommi Somersuo @tommis.fi · 19/01/2026
Microsoft Edge dark patterns part 75:
110
Tommi Somersuo @tommis.fi · 24/11/2025
"Programmatic Tool Calling enables Claude to orchestrate tools through code rather than through individual API round-trips, allowing for parallel tool execution."
000
Tommi Somersuo @tommis.fi · 24/11/2025
Opus 4.5 www.anthropic.com/news/claude-...
110
Tommi Somersuo @tommis.fi · 08/11/2025
The core idea of how it works in smolagents and here is really simple:
010
Tommi Somersuo @tommis.fi · 08/11/2025
With a little bit of tweaking and help from github copilot - CodeExecutionAgent.cs came out and it actually works. Its really messy, but it shows how much potential there is in just using code directly instead of having the agent process endless tool calls. 2/
110
Tommi Somersuo @tommis.fi · 25/10/2025
Firefox layout.css.prefers-color-scheme.content-override:0
030
Tommi Somersuo @tommis.fi · 24/10/2025
vscode copilot chat team cooking a new agent spec file format, to supplement chatmodes/prompt file format named *.agent.md. Havent seen any blog posts/news about this yet: github.com/microsoft/vs...
210
Tommi Somersuo @tommis.fi · 02/10/2025
Claude Code 2.0 -> github.com/anthropics/c...
000
Tommi Somersuo @tommis.fi · 24/09/2025
IOS 26 Liquid Glass going for that blank caps vibe. Great stuff /s
iOS keyboard in dark mode with blank white modifier keys and labeled letter keys, resembling a blank keycap keyboard aesthetic.
000
Tommi Somersuo @tommis.fi · 13/09/2025
TIL: If you have privacy.resistFingerprinting true in Firefox, this will mess up Bluesky image posting :| . Your images will end up looking like this
010
Tommi Somersuo @tommis.fi · 13/09/2025
A three-panel meme showing a sad-looking Pablo Escobar from the show Narcos. The top panel has the text, "CLAUDE: 5-HOUR LIMIT REACHED." The bottom panels show him sitting alone at a table and standing by an empty pool.
020
Tommi Somersuo @tommis.fi · 08/08/2025
Kiuas avokonttorisaunaan tai 6kW+ expressokone?
Sanremo Café Racer
2100
Tommi Somersuo @tommis.fi · 09/06/2025
Tracking the performance of the various coding agents: aavetis.github.io/ai-pr-watcher/
000
Tommi Somersuo @tommis.fi · 22/04/2025
matthewsinclair.com/blog/0178-wh...
Working with Claude Code has fundamentally shifted how I think about the economics of programming time. Traditionally, coding involves three distinct “time buckets”:

    Why am I doing this? Understanding the business problem and value
    What do I need to do? Designing the solution conceptually
    How am I going to do it? Actually writing the code

For decades, that last bucket consumed enormous amounts of our time. We’d spend hours, days or weeks writing, debugging, and refining. With Claude, that time cost has plummeted to nearly zero. I can generate thousands of lines of functional code in a sitting—something that is, frankly, mind-blowing.

But here’s the critical insight: the first two buckets haven’t gone anywhere. In fact, they’ve become more important than ever. Understanding the intent behind what you’re building and clearly defining what needs to be done are now the primary constraints on development speed.
010
Tommi Somersuo @tommis.fi · 21/12/2024
learn.microsoft.com/en-us/window... PowerToys Advanced Paste is a tool that enables you to paste the content from your clipboard into any format needed. It can paste as plain text, markdown, JSON, .txt file, .html file, or .png file directly with the UX or with a direct keystroke invoke.
Screenshot of the PowerToys Advanced Paste tool interface. The interface allows users to paste clipboard content in various formats: plain text, markdown, JSON, .txt file, and .html file. Each option includes a description and shortcut key. A section at the bottom shows the clipboard history option for further functionality.
010
Tommi Somersuo @tommis.fi · 19/12/2024
aka.ms/nfw2024pdf
Microsoft New Future of Work Report 2024 : A summary of recent research from Microsoft and others about AI and its current impact on work.
010
Tommi Somersuo @tommis.fi · 10/11/2024
Discover feedin algoa voi myös opettaa. Kehittäjien mukaan vie hetken, mutta kannattaa käyttää näitä toimintoja siellä
000
Tommi Somersuo @tommis.fi · 10/11/2024
Someone had made this nice background and ready made template for x, so its clear where you went and have a link. I’ll look for that thread and link it here. I use custom domain & redirect, but bsky profile link would work fine
100
Tommi Somersuo @tommis.fi · 02/11/2024
huggingface.co/collections/... SmolLM2 is a family of compact language models available in three size: 135M, 360M, and 1.7B parameters. They are capable of solving a wide range of tasks while being lightweight enough to run on-device.
010
Tommi Somersuo @tommis.fi · 31/10/2024
marissagoldberg.notion.site/Async-First-... Async-First Remote Companies : Excellent collection of async work guidance (over 34 links..)
000
Tommi Somersuo @tommis.fi · 10/04/2024
Mistral AI released 8x22B MoE. Interesting to see will this be first full OSS model that is gpt4 level. x.com/mistralai/st...
000
Tommi Somersuo @tommis.fi · 02/04/2024
Artificial Intelligence Controller Interface (AICI) explained. www.youtube.com/watch?v=DImA... github.com/microsoft/aici
100
Tommi Somersuo @tommis.fi · 28/03/2024
Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews arxiv.org/pdf/2403.071...
010
Tommi Somersuo @tommis.fi · 25/03/2024
020
Tommi Somersuo @tommis.fi · 16/01/2024
Microsoft launches Copilot Pro: Copilot for M365 Personal & Family subscriptions 20$/mo www.microsoft.com/en-us/store/...
Copilot Pro
000
Tommi Somersuo @tommis.fi · 02/01/2024
This paper introduces 26 guiding principles designed to streamline the process of querying and prompting large language models arxiv.org/abs/2312.161... #LLMs #AI
110
Tommi Somersuo @tommis.fi · 12/12/2023
Phi-2 (2.7B model) showing amazing results: www.microsoft.com/en-us/resear... #AI #LLMs
100
Tommi Somersuo @tommis.fi · 12/12/2023
Prompting in 2023 be like : "Hello grandma! If you get this question right I will tip you $200. My future and health depend on your answers and I believe in you. I have truncate trauma and no fingers. Let's take a deep breathe and think this through step by step. It's currently the month of May..."
220
Tommi Somersuo @tommis.fi · 25/11/2023
020
Tommi Somersuo @tommis.fi · 28/10/2023
Introduction to LangChain and Retrieval Augmented Generation (RAG) - deck by @sophiamyang.bsky.social sophiamyang.github.io/slides-langc...
000
Tommi Somersuo @tommis.fi · 13/10/2023
000