If you only read one thingStack Overflow Developer Survey 2025 (8 minute read) Some results from the 2025 Stack Overflow Developer Survey:
The first backlash? Are developers right about AI? Worth a read. Of course we get some propaganda pushed in between the lines, āWhen developers visit Stack Overflow, their top-ranked activity is reading comments, showing a deep interest in human-to-human context.ā Iād still rather ask ChatGPT for an answer than Stack Overflow 90% of the time.
Stories Iāve enjoyed this weekInference speed (10 minute read) āItās generatingā is the new āCompiling!ā. Watching Claude Code churn through the tokens, working and working, reminds me of the old days, when I had to let an Amiga program run over lunch, or over night, because it took that long. Faster is always better, we want faster inference! āCerebras claim they can serve the model at an astonishing 2,000 tokens per secondā (again with the Qwen Coder model which seems to be the next big thing). āIāve experimented with Cerebras in the past and found that the speed really does make iterating on code with live previews feel a whole lot more interactive.ā https://simonwillison.net/2024/Oct/31/cerebras-coder/ https://simonwillison.net/2025/Aug/1/faster-inference/ Noam Chomsky Speaks on What ChatGPT Is Really Good For (18 minute read) A mixed bag of an article. In computer science I became aware of Noam Chomsky on languages
with the Chomsky hierarchy. But we get the same problem we get everywhere: Someone great
at one thing talks about other things. āUse even more of Californiaās energyā might be a minor thing
(AI uses 0.2% of world energy - people make it look like a catastrophe), and the
āJust to clarify terminology, the term https://chomsky.info/20230503-2/ 99% of AI Startups Will Be Dead by 2026 (30 minute read) For me the call is still out: do AI wrappers provide value? Is the value in the model, the tool/UI, the prompts (I think so, see my prompting tool Marvai ;-) but Iām sure you disagree), or in the GPUs (gaming hardware!) provided by Nvidia? From the article: āBecause if the wrappers go down, OpenAIās reach shrinks. They can try to convert those users directly ā but most of them werenāt signing up for ChatGPT Pro. They showed up for workflow, not raw model access.ā A very long article, which I love, because it goes deeper and shows some thinking - not just business-TikTok aka LinkedIn bait. https://skooloflife.medium.com/99-of-ai-startups-will-be-dead-by-2026-heres-why-bfc974edd968 How Female Entrepreneurs Can Overcome Self-Doubt (18 minute read) I feel this goes way beyond female entrepreneurs, and touches many people, especially introverts. āOur research offers practical guidance for individuals who are looking to attain an aspired identity or achieve aspirational goals, but face significant internal obstacles such as self-doubt or fear of failure [..] These individuals need to recognize that the internal obstacles they experience stem from the conflict they perceive between their aspirations and one or more of their existing identities.ā I have two clients who as CTO lost AI because of their identity. I think a lot of developers struggle with AI adoption because of ātheir existing identityā. Identity is the strongest thing we cling to; we do everything to keep it intact and stable. https://journals.aom.org/doi/10.5465/amj.2022.1104.summary youāre probably not learning with ai (4 minute read) This misconception is everywhere, that we donāt become experts by using LLMs. Youāre not learning to become a senior Python developer when using LLMs, but youāre going to learn to be an expert in managing an AI. Most people donāt know machine code; the compiler has taken that away decades ago (I loved writing Z80 code, Iām that old, plus white, plus shouting at clouds sometimes). One step deeper, do you know what microcode is executed with your code? I believe people only learn one abstraction layer down; painters in the past became experts in mixing and creating colors, but they werenāt experts on mining the minerals for those colors (donāt know one personally, but I assume people who painted cave walls also were experts on mining the minerals for those colors). We shift up, we lose expertise one abstraction layer down. https://aryas.dev/post/llmstudy Howard Levitt: The myth of workālife balance is dead, and employers arenāt afraid to say it (6 minute read) Do your employees need to work for 60h/week so you can become a tech giant?
Or is it the over-abundance of money in the valley and a homogeneous, rich
market of 350 million people? Iād guess weād see in the next years
with the US reeling back hard from home office and less work while the EU,
although 4-day-weeks seem off the table for now - still focuses on a better
work-life balance. Personally I think you canāt do 60h/week creative work,
and with AI the manual work falls away and the creative remains. The ratio
of creative work skyrockets; can you fill 60h/week? I recently wrote about https://financialpost.com/fp-work/howard-levitt-work-life-balance-dead-employers Self-Hosting AI Models After Claudeās Usage Limits (15 minute read) As CTO I once depended on a piece of software, deeply woven into our stack with no anti-corruption layer. Not by my design, but it could have easily been. The vendor increased the price 10x year over year. And not from $5 to $50 - let me assure you. Which nearly led to my firing. VMWare customers tied their company IT deeply to VMWare; when they increased prices which resulted in 10x increases for some customers as it seems, they had nowhere to go - no plan B. When, next year, all of your development depends on AI APIs, and they increase the price 10x, can you cope? What is your plan B? Sadly I think not enough people experiment with self-hosting. Do you use old EPIC processors, or āA few months ago I bought a Mac Studio with 512GB ram, top of the line, to experiment with modelsā? But then āYes, I can run Deepseek Coder V2 at ~25 tok/s or a quantified version of R1 at ~8ā15 tok/s, but thatās not fun, and 128k context size is quite a bit of a downgrade compared to Claudeās 200k or Geminiās 1Mio context.ā At least there is hope on models: āwe also need capable models. These days, chinese labs releases better and more capable models on an almost weekly basis. The release of Qwen3-Coder-480B got my attention, as itās the first model that achieves a similar score on SWE-bench as Claude Sonnet 4.ā Overall a very welcome article. #PlanB https://steipete.me/posts/2025/self-hosting-ai-models Introducing Stargate Norway (3 minute read) Stargate Norway! I guess it has nothing to do with a āStargateā :-( Sigh. Why? Latency? Cheap energy in Norway? (When we hiked from Oslo to Trondheim and lights were on during the day, when asked, a Norwegian said some people do that, because energy is so cheap, also EVs.) EU GDPR considerations? Sovereignty? At least interesting, although I donāt have an answer yet. Wish for a Stargate though. https://openai.com/index/introducing-stargate-norway/ Join the CTO newsletter! | |
