JP Beaudry

Of all the podcasts I listen to, @GergelyOrosz at @Pragmatic_Eng has the smoothest ad integration by far. Ads land exactly when they’re relevant to the conversation. I’m sure it’s more work, but I see you. Many otherwise amazing pods could learn from this… cough The Knowledge Project @farnamstreet cough

Super interesting! I tried hard last month to get Claude Code to use a debugger via MCP, and it just wouldn’t. Even banning grep and terminal access didn’t help. It found creative workarounds (e.g. recreated a shell in Python!?) Looks like post-training might be key. Can't wait to read the blog!

I post-trained Qwen3-Coder to fix bugs using an actual debugger. The result: Solve rate: 70% → 89% Median turns to fix: 46 → 19 (-59%) Instead of just reading code or print-debugging, it: - reasons from execution - inspects live variables and call stacks - sets breakpoints, steps, and evaluates expressions

The little things in life: I congratulated Cursor on a great answer

My online passport renewal only took two weeks! Blown away. That's at the routine, non-expedited level, and includes weekends and snail mail shipping. Also, being able to use a digital selfie removed a huge hassle. Well done, @StateDept, well done.

@BattleBots Thanks for the great show! Publish your soundtrack!

Writing excuses over time First - Dictated but not read Then - Sent from my iPhone, please excuse any typos. Now - Hallucination! Next? My agent/robot did it

Reading DeepSeek R1's inner monologue is a jolly good time. "Wait, maybe... I'm getting confused... Wait, I need more context... But it's unclear... But wait, no..." Takes 10 min to replicate with Ollama and the "--verbose" switch.

I wish @OpenAI ChatGPT web would not log me out 1-2 times per day. By contrast, Gemini goes weeks without kicking me out.

@apolloaisafety performed extensive testing and looked at the chain of thought of models like o1. Sure looks like scheming: apolloresearch.ai/research/schem… But @krishnanrohit provides compelling arguments that it's not due to the lack of an "entity": strangeloopcanon.com/p/no-llms-are-… Strange times

Are #LLMs capable of scheming or not? Apollo Research thinks so: "When AI goals are misaligned with humans', weird shit happens" Rohit Krishnan disagrees "because there's no 'they' there" In practice, does intent even matter? Or just answer quality? See their arguments below.

Ref: youtu.be/1WUuppfscXw?t=…

How much AI speedup to expect in software engineering? Not all companies have the same lived experience so far. See this funny take by @danieldines, CEO of UI Path on the @20vcFund pod: Host - Benioff said that software engineer cohorts have been made so much more efficient, they're now using mostly AI tools. How are you seeing efficiency within development teams change internally at UI Path? Daniel - Look, with all respect to Salesforce, our technology is much harder to build. What we are building spans many different technologies. We we will get some productivity improvement but I don't think it'll be gigantic. Kinda mirrors our speedup experience to date.

More details about the book and the author: thinklikeacto.co

Started reading "Think like a CTO" by Alan Williamson out of a sense of duty. So far, it's comprehensive, actionable, and well-articulated! One gem: "Office politics can be defined as the absence of data that is replaced by emotion." Consider it.

References Preso: youtube.com/watch?v=B6PKVZ… OpenHands: github.com/All-Hands-AI/O… SWE bench: swebench.com

Great preso on the state of AI Agents as of early 2025 by Prof Graham Neubig at NeurIPS. I recommend the following personas take a listen • Designers of agents: how the latest research can get continuously integrated • Users of AI coding agents: how the OpenHands open-source agent tops the SWE Bench • Builders of websites and APIs: how structured input/output makes life easier or LLMs and agents

@CARandDRIVER I want one!

This Off-Road-Ready Genesis GV80 Was Made for Racing Legend Jacky Ickx caranddriver.com/news/a63321988…

"All modern software was woven in the loom of sorrow" -Grady Booch That hits close to home

@a16z Got some unique insights about future events? Great place to put your money where your mouth is. A fun informative listen: a16z.simplecast.com/episodes/predi…

Prediction Markets make a lot more sense to me after this surprisingly educational @a16z pod (sometimes salesy). Their accuracy depends on: - # of participants - Sum of money wagered - Diversity of information among participants - Presence of domain experts - Speed and accuracy of resolution