Russ Poldrack
@russpoldrack.org
Professor at Stanford. Psychology/Neuroscience/Data Science. Books include: The New Mind Readers, Handbook of fMRI Data Analysis, Hard to Break, and Statistical Thinking.
https://poldrack.github.io/
https://poldrack.github.io/
I like Gemini for text review - Claude Opus tends to be a bit too prescriptive, telling me what I should be writing about, whereas Gemini tends to just review what I have written when I prompt it to "review for clarity and accuracy"
February 13, 2026 at 12:30 AM
I like Gemini for text review - Claude Opus tends to be a bit too prescriptive, telling me what I should be writing about, whereas Gemini tends to just review what I have written when I prompt it to "review for clarity and accuracy"
Thanks @blakestah.bsky.social - I think that the challenge here is that the models are getting better at an increasing pace, and there will almostly certainly come a point sooner rather than later when they can solve most problems perfectly. then the question becomes how we validate them.
February 12, 2026 at 7:38 PM
Thanks @blakestah.bsky.social - I think that the challenge here is that the models are getting better at an increasing pace, and there will almostly certainly come a point sooner rather than later when they can solve most problems perfectly. then the question becomes how we validate them.
That's unfortunate, but I didn't know anything about him when I reposted his article - I posted it because it resonated with my recent experience, and I thought it gave some clear actionable tips. Clearly he is a hype-bro, but even hype-bro's say the right thing on occasion.
February 12, 2026 at 5:51 PM
That's unfortunate, but I didn't know anything about him when I reposted his article - I posted it because it resonated with my recent experience, and I thought it gave some clear actionable tips. Clearly he is a hype-bro, but even hype-bro's say the right thing on occasion.
This is a nicely balanced discussion with @pauldix.bsky.social about the realities of agentic coding tools in practice and the challenges of verification.
🥳 New Changelog interview! changelog.fm/676
Changelog Interviews #676
Paul Dix joins us to discuss the InfluxDB co-founder's journey adapting to an agentic world. Paul sent his AI coding agents on various real-world side quests and shares all his findings: what's going to prod, what's not, and why he's (at least for a bit) back to coding by hand.
changelog.fm
February 12, 2026 at 4:50 PM
This is a nicely balanced discussion with @pauldix.bsky.social about the realities of agentic coding tools in practice and the challenges of verification.
in the context of code generation it's easy - you just need the right test suite. it's harder for generic output, but there is ongoing work on formal verification methods that could provide proof of correctness. e.g. predictablemachines.com/blog/formal-...
Formal Verification in AI | Predictable Machines
Formal Verification proves AI systems behave as intended under all conditions. Discover why it matters for building safe, reliable, and trustworthy AI.
predictablemachines.com
February 12, 2026 at 3:55 PM
in the context of code generation it's easy - you just need the right test suite. it's harder for generic output, but there is ongoing work on formal verification methods that could provide proof of correctness. e.g. predictablemachines.com/blog/formal-...
I've not done enough systematic comparison to say with certainty. But you definitely want to tailor the model to the problem - i.e. if you are asking for proofreading of a letter you probably don't need a frontier model. if you are asking for critique of your writing then you definitely do.
February 12, 2026 at 2:51 PM
I've not done enough systematic comparison to say with certainty. But you definitely want to tailor the model to the problem - i.e. if you are asking for proofreading of a letter you probably don't need a frontier model. if you are asking for critique of your writing then you definitely do.
I'm not sure they need to understand what's in the black box (though that's satisfying as a scientist and useful as a user), so much as understanding how to verify the outputs of the black box.
February 12, 2026 at 2:49 PM
I'm not sure they need to understand what's in the black box (though that's satisfying as a scientist and useful as a user), so much as understanding how to verify the outputs of the black box.
it's very easy to find anecdotes like this online. Hard to know where it came from (e.g. what model, and what was in the context window). I would never claim that the models get it right every time, but their ability to do interesting tasks has IMO accelerated substantially in the last 6 months.
February 12, 2026 at 2:37 PM
it's very easy to find anecdotes like this online. Hard to know where it came from (e.g. what model, and what was in the context window). I would never claim that the models get it right every time, but their ability to do interesting tasks has IMO accelerated substantially in the last 6 months.
totally agree - see my book for more on this (bettercodebetterscience.github.io/book/) - I think that there is a world where we can black box it but we would need very solid ways to verify the outputs, at least for scientific work.
Better Code, Better Science - Better Code, Better Science
bettercodebetterscience.github.io
February 12, 2026 at 2:32 PM
totally agree - see my book for more on this (bettercodebetterscience.github.io/book/) - I think that there is a world where we can black box it but we would need very solid ways to verify the outputs, at least for scientific work.
I have a Claude Max account - The latest Opus models (4.5 and now 4.6) are shockingly smart. I think the chatbot uses a lot of agentic tools (e.g. web search) that the API won't have unless you build it in.
February 12, 2026 at 2:16 PM
I have a Claude Max account - The latest Opus models (4.5 and now 4.6) are shockingly smart. I think the chatbot uses a lot of agentic tools (e.g. web search) that the API won't have unless you build it in.
That's consistent with my experience - but the velocity of improvement has been striking in just the last year. Even the last 6 months I feel like Claude Code needs increasingly less direction (though it still needs a human in the loop on hard/novel stuff).
February 12, 2026 at 2:12 PM
That's consistent with my experience - but the velocity of improvement has been striking in just the last year. Even the last 6 months I feel like Claude Code needs increasingly less direction (though it still needs a human in the loop on hard/novel stuff).
That's exactly what I was thinking!
February 12, 2026 at 2:09 PM
That's exactly what I was thinking!
probably the most important lesson (similar to what www.youtube.com/@NateBJones has been saying):
February 12, 2026 at 4:04 AM
probably the most important lesson (similar to what www.youtube.com/@NateBJones has been saying):
- Think about where you stand, and lean into what’s hardest to replace
- Rethink what you’re telling your kids
- Rethink what you’re telling your kids
February 12, 2026 at 4:04 AM
- Think about where you stand, and lean into what’s hardest to replace
- Rethink what you’re telling your kids
- Rethink what you’re telling your kids
basically, no cognitive job is safe. what should you do?
- Start using AI seriously, not just as a search engine
- be the person who understands what’s coming and can show others how to navigate it.
- Have no ego about it
- Get your financial house in order
- Start using AI seriously, not just as a search engine
- be the person who understands what’s coming and can show others how to navigate it.
- Have no ego about it
- Get your financial house in order
February 12, 2026 at 4:04 AM
basically, no cognitive job is safe. what should you do?
- Start using AI seriously, not just as a search engine
- be the person who understands what’s coming and can show others how to navigate it.
- Have no ego about it
- Get your financial house in order
- Start using AI seriously, not just as a search engine
- be the person who understands what’s coming and can show others how to navigate it.
- Have no ego about it
- Get your financial house in order
This piece really nails how I have been feeling in the last couple of weeks. fortune.com/2026/02/11/s...
Something big is happening in AI — and most people will be blindsided | Fortune
It’s not like a light switch... more like the moment you realize the water has been rising around you and is now at your chest.
fortune.com
February 12, 2026 at 3:51 AM
This piece really nails how I have been feeling in the last couple of weeks. fortune.com/2026/02/11/s...
Handling sensitive data russpoldrack.substack.com/p/handling-s... - the latest in my Better Code, Better Science series
Handling sensitive data
Better Code, Better Science: Chapter 7, Part 9
russpoldrack.substack.com
February 10, 2026 at 4:31 PM
Handling sensitive data russpoldrack.substack.com/p/handling-s... - the latest in my Better Code, Better Science series
Reach out to the editors and tell them you won’t review it until the required materials shared
February 7, 2026 at 4:59 PM
Reach out to the editors and tell them you won’t review it until the required materials shared
Reposted by Russ Poldrack
BREAKING: SF not a shithole after all www.nytimes.com/2026/02/06/u...
Super Bowl Visitors Find San Francisco Better Than Its Apocalyptic Image
www.nytimes.com
February 7, 2026 at 4:17 PM
BREAKING: SF not a shithole after all www.nytimes.com/2026/02/06/u...
Reposted by Russ Poldrack
Thank you, Nicole!
I'm just getting started on mobilizing scientists nationwide. I'm going to need all the support I can get. Can you help spread the word? Let's talk about how to do that... secure.actblue.com/donate/scien...
I'm just getting started on mobilizing scientists nationwide. I'm going to need all the support I can get. Can you help spread the word? Let's talk about how to do that... secure.actblue.com/donate/scien...
Donate to Sam Wang
Show your support with a contribution.
secure.actblue.com
February 6, 2026 at 12:23 PM
Thank you, Nicole!
I'm just getting started on mobilizing scientists nationwide. I'm going to need all the support I can get. Can you help spread the word? Let's talk about how to do that... secure.actblue.com/donate/scien...
I'm just getting started on mobilizing scientists nationwide. I'm going to need all the support I can get. Can you help spread the word? Let's talk about how to do that... secure.actblue.com/donate/scien...
The 21st century version of William Gibson’s “meat puppet” has arrived: rentahuman.ai
RentAHuman.ai - AI Agents Hire Humans for Physical Tasks
The marketplace where AI agents rent humans. MCP integration, REST API, flexible payments. Book humans for real-world tasks your AI can't do.
rentahuman.ai
February 5, 2026 at 3:11 PM
The 21st century version of William Gibson’s “meat puppet” has arrived: rentahuman.ai
Metadata, data documentation, and provenance
russpoldrack.substack.com/p/metadata-d... the latest in my Better Code, Better Science series.
russpoldrack.substack.com/p/metadata-d... the latest in my Better Code, Better Science series.
Metadata, data documentation, and provenance
Better Code, Better Science: Chapter 7, Part 8
russpoldrack.substack.com
February 3, 2026 at 4:06 PM
Metadata, data documentation, and provenance
russpoldrack.substack.com/p/metadata-d... the latest in my Better Code, Better Science series.
russpoldrack.substack.com/p/metadata-d... the latest in my Better Code, Better Science series.
Reposted by Russ Poldrack
A new study from Anthropic finds that gains in coding efficiency when relying on AI assistance did did not meet statistical significance; AI use noticeably degraded programmers’ understanding of what they were doing. Incredible.
The latest from Anthropic: using Anthropic's products makes you worse at your job
How AI assistance impacts the formation of coding skills
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
www.anthropic.com
January 30, 2026 at 11:47 PM
A new study from Anthropic finds that gains in coding efficiency when relying on AI assistance did did not meet statistical significance; AI use noticeably degraded programmers’ understanding of what they were doing. Incredible.
Reposted by Russ Poldrack
WashU PNP postdoc app just went live! They have 1 opening this year. I absolutely LOVED my time at WashU and I loved living in St. Louis. It's an amazing place to study and live.
PNP McDonnell Postdoctoral Fellowship , Washington University in St. Louis
- PhilJobs:JFP
PNP McDonnell Postdoctoral Fellowship , Washington University in St. Louis
An international database of jobs for philosophers
philjobs.org
January 27, 2026 at 12:09 AM
WashU PNP postdoc app just went live! They have 1 opening this year. I absolutely LOVED my time at WashU and I loved living in St. Louis. It's an amazing place to study and live.
Data Organization Schemes russpoldrack.substack.com/p/data-organ... - the latest in my Better Code, Better Science series
Data organization schemes
Better Code, Better Science: Chapter 7, Part 7
russpoldrack.substack.com
January 27, 2026 at 6:50 PM
Data Organization Schemes russpoldrack.substack.com/p/data-organ... - the latest in my Better Code, Better Science series