Jonathon Bloomfield

@jayvanbavel I wonder how this changes with more recent or intelligent models. I don't think I'd consult any of those models for life advice.

Most participants who had a 20-minute discussion with AI chatbots about health, careers or relationships followed its advice. However, 2-3 weeks later, participants receiving advice from AI showed no sustained well-being. These findings reveal that LLMs exert substantial influence over real-world personal decisions without delivering measurable psychological benefits. arxiv.org/abs/2511.15352

You truly become an adult the day you realise Claude starts its reasoning process with "this is genuinely interesting" in order to force itself to respond as if whatever you just said were genuinely interesting.

@colinarms @conductor_build Thank you so much!

@Champdebloom @conductor_build DaisyDisk. Stellar app with a reasonable one time purchase fee

I somehow ended up with 1TB worth of node_modules in my @conductor_build git worktree folders because every fresh worktree does a `yarn install` What's the solution here? Yarn hardlinks, yarn PnP, move to pnpm, etc?

@evan___yu @ExaAILabs Tragedy :(

@Champdebloom @ExaAILabs you still need to pass in an Exa api key, so it's billed by Exa

We're excited to partner with Google to offer Grounding With Exa inside of Gemini models! Using Exa's agent-first search, Gemini models can now access billions of websites, technical docs, papers, people, companies, and more. 10^18🤝10^100

Five months in, I think I've decided that I don't want to vibecode — I want professionally managed software companies to use AI coding assistance to make more/better/cheaper software products that they sell to me for money.

@codetaur This is what initially drew me to @RepoPrompt last year

one of the most useful things I built lately is just a way for codex to send big ass prompts with the majority of the codebase to gpt-5.5 pro. I still handle the handoff between the browser UI and GPT-5.5 pro but man is it convenient.

@AndyMasley It feels like AI use and understanding compounds

@AndyMasley I think using AI a skill that develops as you use it, in the sense that the people I notice keeping up seem to be more proficient, and notice possibilities those with less experience don’t.

I don't believe at all that people need to "use AI because otherwise they'll fall behind on it." Capabilities are moving so fast that we'll all be novices in 3 years regardless of what we're doing now. My poking around GPT-3 papers gave me basically no leg up on actually using AI compared to people who just dove straight into coding agents.

@paulnovosad It seems to make meetings feel costlier because of how I could've used the time to push forward with LLMs instead.

AI has raised the utility of work so much I'm working about 10h more per week than I used to. The joy of research was always in the exploring and discovering, suddenly I can spend my time doing that again.

@chehendriksen Have you tried @RepoPrompt ? From what I’ve found, this seems like the best way of playing to each model’s strengths.

We like to think about the economics of human-AI collaboration, but with 5.5 and Opus 4.7, I actually use each more because of the other. Opus is better at frontend, so that makes the backend capabilities of 5.5 more valuable. And vice versa.

Introducing Philosophy Bench, my favorite new project I've worked on this year, with help from my friend @matthewjmandel We put frontier language models in 100 ethically complex situations and require them to act, grading them on adherence to consequentialism vs. deontology, tendency to follow user requests, corrigibility, and more 1/

@wesbos @internetofshit I use the offline version of their app, which is supposedly less bloated: hub.sync.logitech.com/options/post/l…

Logitech's servers are down, so LogiOptions+ doesn't load, which means my mouse no longer works. @internetofshit Took 5 mins and ported everything over to Better Touch Tool. Uninstalled that awful app.

@joshgans Maybe that was the plan all along?

When OpenAI completely AI generates one of these livestreams, how will we know given the way the people talk in them?

@joshgans Underspecification seems to be the default for most people, so models asking questions instead of filling in the blanks of their thinking seems like a good call.

Claude Design is cool. What I really like is that it asks you to make a list of choices after your initial prompt.

@Kpaxs This is what makes consulting fun!

The skills that make you a great founder, are the exact things that make you impossible to slot into someone else's org chart. Seeing possibilities nobody else sees, refusing to accept "that's just how it's done", building conviction in your own vision against consensus, tolerating chaos and ambiguity, making decisions with incomplete information, being comfortable with total responsibility.

@pvncher When I hit a doom loop, I ask the model to undo its changes then create a clear description of what we wanted to achieve. Passing this to a new thread or chat tends to produce much better results because of the clarity the model gained in the unsuccessful pass.

I really appreciate how thoughtful factory is about context I haven’t seen anyone refer to context as adversarial before - it’s a great term. I think this actually pushes against endless compactions as a long horizon strategy. You need continuous defragging

@PabloPeniche @TheQuietLynx @signulll Can I try as well? I’d be happy to pay for a month to test.

@TheQuietLynx @signulll I'm happy to give you a discount for you to be the judge! x.com/tenobrus/statu…

these dudes build & ship genuinely great stuff, in this case a fundamental shift in how we interface with our devices. aqua is the best way to input on both your phone & your mac. i have my acqua plugged into my foot pedal at my desk & now i'm using this ios keyboard as a primary input. great tools need to be acknowledged. this is a well crafted tool in every dimension.

@mattrickard This is @RepoPrompt. You can configure your model choices fit different workflows. I have Opus as my agent and 5.4 as my oracle and context gatherers.

Who is building the cross-provider version of this? If it works for Sonnet <> Opus I bet it would work even better if we did multi-executor, multi-advisor Sonnet / Spark <> Opus / GPT 5.4 xhigh