jcheng

7 ms·

31.
by jcheng 10mo ago
That’s exactly what it does, I’ve found it completely un-confuses Claude Sonnet 4.5.
32.
by jcheng 10mo ago
That makes so much sense, it would make a great MCP. Maybe something similar for DOM manipulation; extracting text out of big, noisy HTML pages using a combination of Find Text with selector return values, and a DSL for picking and filterin
33.
by jcheng 10mo ago
I take these early reports (less than a week or two after a major model release) with a grain of salt. It takes time to get to know a model, and maybe there's some selection bias in who's posting within 1-2 days of getting access.
34.
by jcheng 10mo ago
Personally, CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR=1 made all my cd problems go away (which were only really in cmake-based projects to begin with).
35.
by jcheng 10mo ago
Seems like an opportunity for someone
36.
by jcheng 10mo ago
They met tonight! This is so insane!
37.
by jcheng 10mo ago
Reminds me of Stephen Colbert's roast of George W. Bush at the 2006 White House Correspondents' Dinner: > The greatest thing about this man is he's steady. You know where he stands. He believes the same thing Wednesday tha
38.
by jcheng 10mo ago
At the time, Java. J2EE (entity beans and session beans), Java Server Pages, Apache Struts. I think it's hard for people who didn't live through it to appreciate just how painful it was to work in that stack circa 1999-2003, like
39.
by jcheng 10mo ago
Why do you think this is worse than someone saying about Java: "some coworker put `this.` on a variable relating to sessions and suddenly people started seeing each other's accounts"? Because it's less obvious what &quot
40.
by jcheng 10mo ago
Generative, not general
41.
by jcheng 11mo ago
Interesting that they run them upside down.
42.
by jcheng 11mo ago
Automatic speech recognition and speech to text models are also growing up real fast.
43.
by jcheng 11mo ago
Wow, it's true--Terminal is <canvas>, while the editor is DOM elements (for now). I'm impressed that I use both every day and never noticed any difference.
44.
by jcheng 11mo ago
Sure, I’m with you on your larger point.
45.
by jcheng 11mo ago
Do you have a citation for that factoid?
46.
by jcheng 11mo ago
Seems almost quaint in late 2025 to object to a workable technique because it "seemed like a bit of a hack"!
47.
by jcheng 11mo ago
AFAIK, every tool out there that lets you do oauth with your Claude Max plan is doing so with the same copy-pasted client id/secret that are extracted from Claude Code. It's not clear at all that this is above board, and when I as
48.
by jcheng 11mo ago
Slightly off topic, but I wonder if Quarto would work for you ( https://quarto.org , disclaimer, I work for the company that maintains it). You write Markdown files with code blocks, and they're rendered into nice looking doc
49.
by jcheng 11mo ago
512MB file of incredibly compressible data, then?
50.
by jcheng 11mo ago
It's not quite as trivial as that; one could start the page with a <script> tag that contains "", and that would hide all the content from your scraper but not from real browse
51.
by jcheng 11mo ago
The layoffs at Amazon and Microsoft are not due to lack of profits. They’re massively profitable right now. https://www.macrotrends.net/stocks/charts/MSFT/microsoft/ebi... https://www.macrotre
52.
by jcheng 1y ago
Not if you're passing binary data
53.
by jcheng 1y ago
Why taxpayers? Where’s the systemic risk in AI labs getting acquired for cents on the dollar? The taxpayers weren’t holding the bag during the dot com crash, just investors.
54.
by jcheng 1y ago
For 2, a lot of companies use AWS Bedrock to access Claude models instead of Anthropic, for exactly this reason. Amazon’s terms say they don’t log prompts or completions and don’t send anything to the model provider. If your production data
55.
by jcheng 1y ago
No, I didn’t provide any tools
56.
by jcheng 1y ago
The difference in coding ability between then and now is pretty huge. And a year ago o1 hadn’t been introduced yet, whereas now the “reasoning” technique is pretty widespread. Not sure if you’re counting things built on top of the models bu
57.
by jcheng 1y ago
I want this for after the code has run and returned results. Often when you use code to answer questions about a table, the result is in the form of a smaller table. I'd like to know how small that table needs to be before you can rely
58.
by jcheng 1y ago
uv add google-genai uv run scripts/run_benchmarks.py --models google/gemini-2.5-pro --formats markdown_kv --limit 100 And add GOOGLE_API_KEY=<your-key-here> to a file called .env in the repo root. Unfortunately
59.
by jcheng 1y ago
gpt-5 also got 100/100 for both CSV and JSON. uv run inspect eval evals/table_formats_eval.py@table_formats_csv --model openai/gpt-5 --limit 100 uv run inspect eval evals/table_formats_eval.py@table_formats_jso
60.
by jcheng 1y ago
I was curious enough to have Codex create a similar benchmark: https://github.com/jcheng5/table-formats With 1000 rows and 100 samples and markdown-kv, I got these scores: - gpt-4.1-nano: 52% - gpt-4.1-mini: 72% - gpt-

More ›