Product guy in health tech. Ex-founder. Starting to build again.
New York, NY
Joined October 2009
- Tweets458
- Following195
- Followers47
- Likes110
using instinct through imessage is nice, feels like texting a person. but there's such a thing as too simple. asked it to book a table, silence, then a verification code out of nowhere. worked fine this time. with muse you can at least watch it happen.
what makes muse a good daily driver is the default tone. short, friendly, straight to the point. with claude or chatgpt you'd have to set that up yourself.
we all watched claude code have its moment with enterprise. now it feels like muse is about to have that same moment with consumers.
Chris retweeted
One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.
In the world of AI agents, I’m realizing that understanding why a metric is going up unexpectedly is just as important as understanding why it’s going down.
Example: today, our agent containment rate increased by ~10 percentage points without any changes on our end.
The easy thing would have been to chalk it up as a win. But after digging in, we realized the increase was actually caused by something that negatively impacted the end user experience.
It’s a good reminder that with AI agents, higher containment is not always better. Metrics can move in the “right” direction for the wrong reason.
Best quote I’ve heard to describe AI- it’s alien technology with no instruction manual. Really resonates…not the alien tech part, but the “no instruction manual” part. Shout out to Karpathy
Chris retweeted
Morgan Stanley has again raised its capex forecasts for the five hyperscalers Amazon, Alphabet, Meta, Microsoft, and Oracle. It now expects them to spend about $805bn this year, up from a previous estimate of $765bn. For next year, the forecast has been lifted from $951bn to $1.1TRILLION. To put that into perspective, their 2026 spending alone would be roughly equal to what all non-tech companies in the S&P 500 spent combined in 2025. The expected ~$800bn for 2026 is nearly double 2025 levels and about three times what was spent in 2024.
I’ve never been a big fan of social media - don’t really use Instagram, Facebook, LinkedIn, or X (up until 2 weeks ago). But just because I don’t use it / like it doesn’t mean it’s not the future.
All the “eyeballs” are on social media now - it’s just not social anymore…it’s the primary source of information. Is there a new term for this?
Has anyone successfully built an “On-call diagnostic agent” that runs through an on-call SOP and generates a list of potential hypotheses even BEFORE the on-call engineer gets online?
With traditional search, different questions often led to the same result. (Ex: Google)
With AI, the wording matters immensely. Change the prompt and you’re not just refining an answer, but you’re potentially exploring a completely different branch of content.
One thing that I’m learning about X. Many of the same ideas are reiterated. But users articulate the same ideas in different ways that different users are drawn to.
Pretty wild that “AGI” was coined mainly to distinguish itself from narrow AI.. primarily a reaction to separate itself from “ANI.”
if we can’t define AGI precisely, how will we ever agree we’ve reached it?
Under appreciated rollouts of Voice AI - Wendy’s drive through. Pretty cool to see all the other use cases where this is popping up. It’s such an obvious application once you see it.
My 2026 prediction:
AI labs will stop over optimizing on existing benchmark scores and start inventing new ways to measure “real” model quality.
I built a trade anomaly detector for Polymarket about 4 months ago (I forget which model). Wasn’t great, and I paused on improving it
This weekend, I rebuilt the entire thing with Opus 4.5 and got way better results. Crazy how easy it is to “throw away and rebuild” these days
Thoughts from my first 10 days actually using X - Sentiment on X is a more valuable indicator of a model’s adoption (GPT 5.2, Opus 4.5) than benchmark performance.
Not sure if it’s 100% true, but I’ll say one thing - I haven’t heard a positive thing about Microsoft Copilot yet
news.ycombinator.com/item?id…
Chris retweeted
Amazing! @NotebookLM just rocks!
New @NotebookLM: The Collected Works Of Jane Austen. Full text of all of her novels, letters, juvenilia, bio, and more. We designed this one for fan-fiction creators as much as for students and scholars. notebooklm.google.com/notebo…
The Austen Featured Notebook is a particularly compelling showcase of our new Slide Deck feature. We've got everything from a sophisticated overview of the underlying economics of Austen's plots, to a biographical account of how she effectively invented a new career path of a commercial female novelist -- all the way to a proposed graphic novel treatment for a Pride and Prejudice sequel where Mr. Darcy turns out to be a vampire.
Everyone’s trying to measure “dev productivity" from AI tools like Opus 4.5, GPT 5.2
Did we do this for Git, IDEs, or cloud services? Kind of just felt friction disappear and shipped differently.
*And yes, I added Opus 4.5, and GPT 5.2 unnecessarily above to test the algo ;)
