People make such a big mistake thinking that ai useful because it makes errors. Imagine if we had the same thought process with humans!

Perplexity deep research is deep on sources but shallow on response. Still very useful!

The technology overhang will grow until absorption improves

My wife has so many more ideas for deep research than me. It’s so over…

Starting to feel the AGI

The year is 2025. ChatGPT Operator exists. I still surf the web manually...

Deep Research just absolutely nailed the itinerary for an overseas trip my family is planning. I gave it a ton of parameters and it just knocked it out of the park. Would have been hours of work for me at least

ChatGPT Operator is going to need A LOT of access to personal information to be useful for a killer use case. Dealing w/ government

Man Claude Sonnet, how is it still so good. For our testing in calculating complex fields with leases it’s right up there with o3-mini. GPT-4o in the dust…

I think a lot of people are going to have the Oh Shit moment with Deep Research like programmers did with tools like Cursor. This stuff is powerful

It’s highly rewarding to be a news junkie working in AI. Except for…all the other reasons being a news junkie is not good right now

Confession: my cat was sitting on me and I moved before she was ready.

OpenAI Deep Research feels like another GPT4 moment for me. Its ability to carry out complex instructions is unbelievable.

This. This is the way to build apps now

Jevons paradox but for software engineers…

Why is genmoji the best part of Apple intelligence?

I am humbled to be offered to opportunity to pay $1000 to be named a visionary proptech leader to watch. Truly a watershed moment for my career…

If you use your LinkedIn connection request to try to sell me something I will block you. At least have the decency to spend an InMail credit.

I read the DeepSeek R1 paper and so should you. They used RL and set the model loose on verifiable coding tasks, among other things. arxiv.org/abs/2501.12948

www.librechat.ai This thing is incredibly feature rich. Why have I never seen anyone talk about it before?

Out of credits for the month on Cursor. Good thing Windsurf is somehow the same product? How did that happen?

Perplexity will use wolfram alpha for calculations. This is cool and useful! Helped it ace a question ChatGPT got wrong. If I contribute x dollars in my paycheck, when will I hit the 2025 401k max?

Ai rocks at infrastructure as code. As a result I can spin up a poor man’s version of so many saas products I’d otherwise have to spend lots of money on

“The future of AI competition will be about 'power dominance' - do you have access to enough electricity to power the datacenters used for increasingly large-scale training runs” eta-publications.lbl.gov/sites/defaul...

This is helpful on the changing nature of software engineering with AI

I’m glad Cursor was broken for a week because it gave me a reason to test Windsurf. Which is maybe more awesome?

Big not influx to Bluesky happening now?

I wrote about what was important in a crazy month of AI advances: 1) Intelligence, of a sort, everywhere 2) The first very smart models that can push us to push the boundaries of knowledge 3) AI got eyes to go with its ears 4) Leaps in video generation www.oneusefulthing.org/p/what-just-...

Good news for your end-of-year reading list: We've just released the Microsoft 2024 New Future of Work Report! It's packed with insights on how AI is transforming work. aka.ms/nfw2024 #NewFutureOfWork #NFW2024 #AIResearch #WorkplaceInnovation #GenAI

GitHub CoPilot is free now. Cool! Try it out...have to search for the "multi edits" feature. It stinks. Why exactly have they not copied Codeium or Cursor? Or hell bought them?

Wow Cursor absolutely nuked their product. deleting code randomly, going in circles. Good thing Codeium’s Windsurf Agent has the same magic that Cursor had

Is Apple Intelligence Apple Maps part 2? Panned on launch, excellent years later, but overall a missed opportunity?

GPT4o-mini is really fast and smart. Starting to understand why they are calling these models agenetic

It takes a little prompt engineering to unlock the value, but Gemini Deep Research is VERY VERY interesting.

I'm starting to do pull request style reviews with Cursor after a batch of changes.

I have now run enough o1 experiments with fellow academics on their hard problems in fields ranging from neurology to genetics to economics that I believe that it has genuine potential to help with science & academics should try it to see how it does for them. Even when it is off, it spurs thinking

I agree with Kurt Gödel in the Oppenheimer movie. Trees really are the most inspiring structures.

Having Cursor Agent go back and check previous git commits to see where we messed up in a refactor

Gemini code execution is a nice new feature but it’s super limited compared to code interpreter on ChatGPT. No packages, no IO. Basically it’s a fancy calculator

Regret to say that Gemini 2.0 spatial understanding does not accurately on satellite images

It’s too bad cashmere socks are wildly impractical because wow

Sora is such a disappointment. All that wait and it’s worse than Luna on initial launch? Am I missing something?

O1 can demonstrate a trie data structure in ascii, which is a thing that OG GPT4 couldn’t do. GPT4o gets the explanation right but the ascii wrong.

Fed researchers fed Fed minutes from 2010-24 into AI tool to see how it processed the content of discussions. "We were especially struck by how well the models performed despite the fact that we employed them off-the-shelf with no domain-specific fine-tuning" www.federalreserve.gov/econres/note...

One of the most important debates in tech right now is the group of folks who think AI is fake and sucks vs. the people who think AI is real and dangerous. I wrote about why I'm in the latter camp, and talked about my differences with Gary Marcus www.platformer.news/ai-skeptics-...

10 seconds to implement a feature with AI. I am not making this stuff up! It actually works!

To err is to be AI. Or was it human? I forgot.

I think firms worrying about AI hallucination should consider some questions: 1) How vital is 100% accuracy on a task? 2) How accurate is AI? 3) How accurate is the human who would do it? 4) How do you know 2 & 3? 5) How do you deal with the fact that humans are not 100%? Not all tasks are the same.

I asked the Cursor Agent to explore the structure of a word doc I loaded with Python. It tried 6 different approaches on its own. This took about a minute. Then it updated the code. It worked. 👨‍🍳