@kerfors.bsky.social
Lifelong Learner and Information Architect caring about clinical trial data and metadata
I wrote about the connection between two big recent stories in AI: the sudden increase in sycophancy from ChatGPT and the growing evidence that AIs can be hyper persuasive.
First we shape our tools, then they shape us. www.oneusefulthing.org/p/personalit...
"If you want to prompt ChatGPT 40 times, you can just stop your shower 1 second early."
"If I choose not to take a flight to Europe, I save 3,500,000 ChatGPT searches. this is like stopping more than 7 people from searching ChatGPT for their entire lives."
If someone tells you that coding with LLMs is easy they are (probably unintentionally) misleading you. They may well have stumbled on to patterns that work, but those patterns do not come naturally to everyone. Iβve been getting great results out of LLMs for code for over two years now. Hereβs my attempt at transferring some of that experience and intution to you. Set reasonable expectations Account for training cut-off dates Context is king Ask them for options Tell them exactly what to do You have to test what it writes! Remember itβs a conversation Use tools that can run the code for you Vibe-coding is a great way to learn A detailed example Be ready for the human to take over The biggest advantage is speed of development LLMs amplify existing expertise Bonus: answering questions about codebases
Here's the table of contents for my lengthy new piece on how I use LLMs to help me write code simonwillison.net/2025/Mar/11/...
11.03.2025 14:11 β π 338 π 67 π¬ 14 π 15
Fellow NSA - National Security Agency veterans. Look at whatβs happened at the National Cryptologic Museum. They covered up with brown paper the photos of Women in American Cryptology. All in response to President Trumpβs anti-diversity executive order.
02.02.2025 04:51 β π 12504 π 6739 π¬ 935 π 1428
See my recent LinkedIn post www.linkedin.com/feed/update/...
15.01.2025 10:29 β π 0 π 0 π¬ 0 π 0
PDF table extraction tools, mainly Docling/Tableformer and Camelot. Camelot is what I ended up with. Main limitations:
- Best results for simple table structures
- May not handle complex multi-page tables well
- May include footnotes that need manual separation
- Some protocols may fail extraction
When will the true PDF API functionality be available? I was so happy to see the news yesterday and so disappointed this morning
18.12.2024 21:45 β π 0 π 0 π¬ 0 π 0
This morning I manage to do what I have envisioned for some time: I used @anthropic.com's MCP to write files from Claude Chat app to my file directory. I managed thanks to All About AI (Kristian Fagerlie's) great video. However, to really scale this need to be made simpler for end users.
12.12.2024 16:13 β π 1 π 0 π¬ 0 π 0
I really like the concept of: AI and "an over-eager intern" so what has perplexity to say about :-)
27.11.2024 07:06 β π 1 π 0 π¬ 0 π 0
....
24.11.2024 21:36 β π 0 π 0 π¬ 0 π 0
Part of my the first chat with Claude 3.5 Sonnet that I marked with a star. As it maid me both smile and scream :-) Claude "got carried away" and forgot about it's own capabilities.
24.11.2024 20:26 β π 0 π 0 π¬ 1 π 0
