Neopets just arrived in Fortnite, which felt like a good time to go public with some of my related but greyer engineering history. Happy to answer any questions anyone might have!
Perhaps if less spending went towards their private aviation interests LTT labs could review a piece of hardware that was released _this_ year, or maybe extend their narrow testing process to cover real-world use metrics like TTFT. Not to mention the lack of real value-perf comparison to CUDA
As someone who's self-taught, I find this argument tantalizingly validating, but I'm constantly reminded that universities offer more than just the material. Between peer networks and "alumni" connections or even just the ability to email a former professor or TA for advice, these opportunities just aren't really available independently. Skipping college might not limit what you can learn but it sure narrows the network and institutional support available to you.
A good start. From 2024: "A company allegedly tracked people’s visits to nearly 600 Planned Parenthood locations across 48 states and provided that data for one of the largest anti-abortion ad campaigns in the nation, according to an investigation" - https://www.politico.com/news/2024/02/13/planned-parenthood-...
Why? The exfiltration vector was known, the sample size was small, and the safety instructions were likely statically positioned. In regular operating practice, none of these three guarantees may hold.
This should be an indication of how valuably xAI sees the training data that Cursor has accumulated, especially with its work on Composer 2.5. Last month, SpaceX and Cursor announced that they had been working on training a new model from scratch. Interested to see if this will put Cursor back into the zeitgeist.
Yak-shaving-shaming puts limits on the creativity of talented engineers by constraining them to existing patterns and practices or building on top of abstractions, and practically, that results in engineers and teams with less breadth. In an applied software world that's exploded in framework and library complexity in recent years, I think there are always going to be yaks in dire need of a shave.
Yes, exactly. Over the past few months I've been trying to refine the vision here - it started with me thinking about my own use case in personal AI, and now has grown into me understanding how there's actually an pervasive governance problem in agentic applied AI right now - we know inference is inherently fallible and yet have none of the open standards for trust and safety we know can protect consumers and businesses.
For fun, this month I shipped Foreign Trivia - a wordle-style tiny daily trivia game where every question is in a different language, with tap-to-translate words as you play. It's part language-learning, part pattern-matching and part trivia. A new puzzle every day. https://foreigntrivia.com
I've also been working on an open source protocol / reference implementation for user-owned AI memory, with the basic idea being that as applied AI scales, more products will derive more claims about users, teams and workflows from chats, docs, calendars, emails, etc, and they shouldn't be trapped inside of one product. Right now there's a lot of opinions on what shape memory should take internally, but I'm focused less on standardizing that part, and more focused on the primitives around it: requiring inspection, correction, revocation and treating portability as first-class. It's early but I'm starting to build more of a clear vision around it and would love feedback from anyone working on local-first software, personal data stores, capability security, knowledge graphs, etc. https://github.com/danielrmay/likewise
I hope the news moves this debate past "open weights vs. closed APIs" as the only axis. Open weights matter, definitely, but applied AI also needs open infrastructure around the model and it feels a bit like I'm yelling into the abyss highlighting the future we're incentivizing - cognition rented from a few institutions with access changing based on policy, geopolitics and platform incentives like advertising
Thank you! I'm actually using this adversarially (or maybe just an experiment) to compare it with an open source protocol that I've already begun to publish but am having trouble getting traction or review on due to it sitting awkwardly across so many domains: local first, capability, ux, security, personal AI memory, knowledge graphs. I have a reference implementation in Rust (I see this site built theirs in python - interesting) but I've been working more on building the right way to explain the need.
It's hard to explain briefly, and so putting this prompt up was a way for me to possibly generate some interest and act as a little public marker for an idea: open-source user-owned memory infrastructure for AI and the importance that I think it represents. My vision and belief behind this project has been slowly building for the past two months - I think personal AI memory will become one of the most important layers in computing, and I'd like that layer to be inspectable, correctable, portable and truly owned by the humans it describes. I'd like to encourage any casual readers who might be interested to reach out to me.
I've experienced this too - it's as if the security classifiers aren't keeping up with model intelligence. I'll leave the implication of that to the reader.