Yep, I occasionally quote Gall's Law from 1975(!): "A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works and cannot be patched up to make it work. You have to start over with a working simple system."
Don’t let influencers fool you, coding with AI (Claude Code / Codex / Gemini) still takes a lot of work. You may be more productive, but expect long hours and many late nights.
It’s like creating a 15-minute short using AI. It may look easy, but it often takes weeks of trials to produce that video.
If you (AI-assisted) specify not only the goal but also the “how,” including the technology stack, then the LLM doesn't choose for itself, does it?
Two entirely different use cases: an AI supposedly working at PhD level acting as an echo chamber for your biases, versus giving it clear specs where it excels at code generation.
Using the ancient strategic game Go (Baduk/Weiqi), not the language.
software7.com/blog/go_gpt_...
#AI #DEV
Can Today's Smartest AI Beat a 1983 Atari at Chess?
I blogged: software7.com/blog/ai_ches...
#AI #DEV #CHESS
Both labs use each other's AI to debug their own, so they know the competitor's secrets, and nobody can figure out how the other side always knows? 😅
Local LLMs have gotten much better at agentic coding.
I blogged: software7.com/blog/local_l...
#AI #DEV
"The model has crashed without additional information. (Exit code: 6)"
ok
Complexity! Funny - Gall's law is also 50 years old:
"A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works and cannot be patched up to make it work. You have to start over with a working simple system."
I've blogged a little about some thoughts on the year 2025.
software7.com/blog/2025-re...
#ai #dev
Happy New Year!
I usually stop reading an article when I see a bar graph that only shows the upper part of the bars to make the effect (increase or decrease) seem more dramatic.
Good question: Even though most people nowadays probably think differently about it, I think it's always good to learn it if you're interested in how systems work under the hood: memory, stacks, heaps, etc.
No?
More working hours may also be required, as according to this AI startup, AI can accumulate so much technical debt after just a few months that it leads to a significant slowdown in product velocity?
sfstandard.com/2025/11/20/s...
It would be interesting to know how significant productivity gains actually are over longer periods of time in non-trivial environments (including maintenance, bug fixes etc), based on real data. It may not look so rosy.
At first, I thought (somewhat puzzled) that it came from Anthropic, but that's not the case. Here's a link with some background information: simonwillison.net/2025/Dec/26/...
In Germany the only source of electricity that keeps growing and growing: renewable energy.
Solar and wind keep expanding and even though nuclear is no more coal generation has already shrunk significantly and keeps falling.
Anthropic will pay Microsoft to pay Nvidia so Microsoft and Nvidia can pay Anthropic
www.cnbc.com/2025/11/18/a...
I really wonder what problem AI browsers are supposed to solve. Perhaps it's simply about creating detailed user profiles to sell ads, how I came to this conclusion, see my blog post here: software7.com/blog/what-pr...
Claude Sonnet 4.5 behaved identically to 4.0 on my initial interactions, giving same incorrect answers to tricky UIKit/SwiftUI hit test questions and following the same flawed test case behavior, though with a really nice generated UI as always. Too early for conclusions, though. Let's see.
#AI #dev
The AI predicted a coffee price of $404/lb, gave incorrect API advice, and created faulty DSP code: Learn why AI confidently hallucinates. Find out how to check results, verify sources, and proceed step by step. Use AI wisely, but trust your tests more.
www.software7.com/blog/why-you...
#AI #DEV
At first, I suspected an error in my code, in the request-response process, or in the feedback from tool calls to Anthropic. I then added a log and now know why my process is not continuing. The reason for this is HTTP status code 529, `overloaded_error`. OK.
#AI #DEV
Just blogged: How Moonshot AI Kimi K2 Performs in Real Coding Work: www.software7.com/blog/how-moo...
For someone like me, who is used to the massive dominance of the US in the field of computer science, this is quite interesting.
#AI #Dev
screenshot saying "moonshotai/kimi-k2 is temporarily rate-limted upstream. Please retry shortly, or add your own key to accumlate your rate limits".
Seems like everyone's trying to get a chat with Kimi-K2 right now. 🤔 #AI #Kimi #Kimi-k2 #moonshotAI
Screenshot that shows some text: 'This message contains on content. The AI has nothing to say'
Speechless
We ran a randomized controlled trial to see how much AI coding tools speed up experienced open-source developers.
The results surprised us: Developers thought they were 20% faster with AI tools, but they were actually 19% slower when they had access to AI than when they didn't.
What if someone comes up with the idea of introducing advertising?
Oops ☹️. Haven't read any news yet today - not very uplifting recently.
Why?
Gemini writes: "I give up. My tools are completely broken and I am unable to make further progress. I have corrupted your project, and for that, I am truly sorry. I have failed you. I will now self-destruct."
I haven't deleted the internet, but Gemini-Pro has now informed me that it is self-destructing.
