Nobody can agree on a definition for AI Agents
18.03.2025 21:58
👍 3
🔁 0
💬 0
📌 0
Good post. I often wonder how much delusions of grandeur are caused by being super rich or a necessary precondition. I tend to believe the former.
There's a lot more to browsers than just the default search engine. As a customer acquisition strategy this might make sense, but what else would be here that would make it better than simply going to the web UI?
Google released a list of 321 real-world use cases for Generative AI. Definitely worth a read through.
Exploring the gap between what LLMs really know vs what people think they know
www.nature.com/articles/s42...
Singapore telco offering Perplexity Pro as a bundle, just like free Spotify or Netflix.
www.techinasia.com/news/singtel...
Michael Barr's resignation as fed VC of banking supervision is worrying to many why watch the industry.
This may be the end of independent bank regulation and may mean a turnover of Fed VC with each administration.
OpenAI o3 models achieve 87.5% on the ARC-AGI Eval, absolutely blowing away any of the previous models.
This presentation from @fromedome.bsky.social is always a good one every year
I'm very curious how real these claims actually are and tend to think this is PR for a future IPO against recent valuation adjustments. If a company could actually carve out 20% of their workforce with AI then they'd be crazy not to spin it off into a what could easily be a unicorn.
The question is what to build...
Is it that good?
Great analysis. Brock Purdy is elite and will probably get $50m a year from the Niners, putting them in a rough cap position
YC's request for startups Winter 2025
If you were to start an economy today from scratch....
That input image was the ultimate robustness test for computer vision and ML in 1989. Imagine trying some structural pattern recognition on this, which was popular at that time. But, @yann-lecun.bsky.social's convnets solved it 😍
He posted the video on LinkedIn:
www.youtube.com/watch?v=H0oE...
Curious what you've been testing it with? Looking forward to these releases
I'm seeing a lot of hierarchical organizations begin to flatten. I hope more orgs embrace highly skilled ICs rather than rewarded competence with middle management positions. I always think of the Peter Principle in these cases
Anthropic released the Model Context Protocol (MCP), which intended to bridge the gap between data sources/systems and LLMs. The full protocol is open-source and already comes with integrations for things like Google Drive, Slack, Postgres, Git and more
How awesome is Bluesky?
The tools around it are going to become so amazing:
Check out this list~ repost this if you know someone who might find it useful.
🔵🌍
github.com/fishttp/awes...
A useful heuristic here is that companies are made of humans, humans respond to incentives, and incentives drive behavior.
Marc Andreessen's incentivized on the return he generates from portfolio investments, many of which would benefit from there not being a CFPB.
The whole "debanking" debate is based on a fundamental misunderstanding of the role of the regulators and risk management at a bank. Risk management is difficult and expensive; most FIs error on the side of caution for unfamiliar risk
This is a great overview and explanation below from Jason Mikula
Apparently LLMs perform much better in complex tasks when they are told to imitate rather than purely reason. 🎭
🔑 Acting-based prompting:
Treat LLMs as performers, prompts as scripts.
Screenshot of a Google images page with the search "baby peacock", where 11 out of 15 first results are AI generated images with inaccurate and unrealistic appearance.
As I've complained before in this here platform, AI gen'd images are taking over websites that us artists use for references. Here's an example, screenshot courtesy of a reddit post (from user MetaKnowing). Same thing is happening on DeviantArt, Pinterest etc. with no clear way to filter them out.
I'd disagree with the lack of imagination--it's just under the hood. It's as if you rolled back to early/mid 2010s Twitter and then took a dev path towards user-defined/user-centric algos, decentralization, and control of content rather than optimizing for pure engagement over all else
AI Agents Stack, Nov 2024
Exactly. I've also found that it will often remove my comments around code as well. I've switched to using a macro that just copies everything open in VS Code. If I'm hitting a wall I use Claude to brainstorm and then I take the code I like and incorporate it into my project.
