korrents
Thorsten Ball

What Thorsten Ball thinks about LLMs

@thorsten-ball · 70 positions · 2 changes of mind · On ppll ↗

Co-founder of Amp and co-host of the Raising an Agent podcast. Wrote Writing an Interpreter in Go.

Thorsten Ball did not write this page.

We collected these quotes from things they published elsewhere, and every quote links to where it was said. They have no account here and have not endorsed this site. Quotes are word for word; the short line under each one is our own restatement, not their wording. Their own site. Is this you? Claim it or ask us to remove it. Or tell us what is wrong here.

Filter LLMs · Newest first
Topic
All 70AI agents 13software quality 11LLMs 10developer tools 9coding agents 8simplicity 5agency 4design 4AI and human skill 3
24 more topics
AI and jobs 3innovation 3Ruby 3TypeScript 3vibe coding 3writing 3ambition 2optimism 2remote work 2software performance 2team size 2AI slop 1code generation 1compilers 1documentation 1Europe 1Google 1humility 1infrastructure 1MCP 1open source 1patience 1product management 1startups 1
Sort
NewestOldest

Only Thorsten Ball on LLMs. See everyone on LLMs

10 dated positions, 2025 to 2026, in their own words. Our reading of what Thorsten Ball has said — not written or endorsed by them.

  1. Our readingAll indications point towards price/intelligence dropping.

    Their wordsBut all indications point towards price/intelligence dropping.

    ↗@thorstenball on Xx.com

  2. 3 days earlier

  3. Our readingUse the latest and best models: if a model is wrong 10% of the time you cannot tell which 10%, so a cheaper one costs you extra attempts.

    Their wordsUse the latest and greatest models. Everything else is potentially a waste of time. If a model is wrong 10% of the time, you don't know which 10%, so you end up making multiple attempts.

    ↗How I work with Agentsx.com 1st of 14 in this piece

  4. 3 weeks earlier

  5. Our readingViews about AI models formed 4-6 months ago should be reconsidered.

    Their wordsTo repeat: reconsider everything you thought about these models 4-6 months ago.

    ↗@thorstenball on Xx.com

  7. Our readingThe biggest source of confusion about whether AI models are good or bad is that people judge them from wildly different codebases, languages, teams, and usage levels.

    Their wordsBiggest source of confusion is that everyone - spent wildly different # of tokens in last 2 years - works in diff codebases - with diff languages - and diff requirements - in diff teams - with diff processes - in diff companies and then says "models are [good|bad]"

    ↗@thorstenball on Xx.com

  8. 7 days earlier

  9. Our readingAI models are now much more than text-to-code converters.

    Their wordsThese models are so much more now than text-to-code converters.

    ↗@thorstenball on Xx.com

  10. 5 days earlier

  11. Our readingThe terminal is dead; most developer tooling including shells, text editors and CLIs will be washed away by tokens and no longer used by humans.

    Their wordsThe terminal is dead. Most developer tooling will be washed away by tokens. Shells, text editors, CLI tools won’t be used by humans anymore. There’s no need to know command line flags and jq invocations anymore.

    ↗What I believe about the future of software developmentthorstenball.com 6th of 15 in this piece

  13. Our readingTokens are the new computing paradigm on which everything will be remade as computing enters a post-binary era.

    Their wordsTokens are the new computing paradigm. Everything will be re-made on top of it. We’ve had deterministic computers for 80 years, so we confuse “how computers have worked” with “how computers must work.” We’re entering the post-binary era.

    ↗What I believe about the future of software developmentthorstenball.com 7th of 15 in this piece

  15. Our readingSome people will be priced out of producing software for the next few years if they cannot get enough tokens, putting them in a second league.

    Their wordsYes. Some people will be priced out of producing software. Just like only some people could afford a personal computer in the 80s and 90s, for the next few years, if you can’t get enough tokens, you’re playing second league.

    ↗What I believe about the future of software developmentthorstenball.com 12th of 15 in this piece

  17. Our readingAI models considered smart today will be considered very dumb in the future, and smarter models make fewer mistakes and need fewer turns.

    Their wordsAnd what we consider a smart model today will be considered very dumb in the future. But a smarter model makes less mistakes, needs fewer turns.

    ↗What I believe about the future of software developmentthorstenball.com 13th of 15 in this piece

  18. 16 months earlier

  19. Our reading · no longer heldProgramming with LLMs is like playing a slot machine: whether you win comes down to luck, so effort spent on prompting is wasted.

    Recalls holdingIt’s not a slot machine and it’s not a coin toss when you ask it to complete something.

    SaidFine-tuning showed me that effort, and care, and patience are rewarded.

    ↗There's Beauty in AIregisterspill.thorstenball.com