Anthropic and OpenAI have certainly been working on this behind the scenes, while they try to see how much better they can get models, they will let others pay for the current state until they find it valuable. The shift we are seeing now is already happening, and they are taking an even larger macroscopic approach by creating computer/tool use, along with the context protocol, so that when it's released it will work with almost any IDE and system...
Similar experience. Whenever I have questions where I know what answer should be (i.e. errors, how to do something etc, ELI5) I will use LLMs. When I use google now I almost never use the AI overview because I am almost exclusively trying to discover something that I want from a source. There's also a whole interesting category of search called navigational queries where there is low entropy (and thus ad value) when someone just goes to google to search espn.com, because they will almost always not explore. I do not understand why people would use perplexity or anything over just chatgpt,claude etc unless some price thing - because it occupies this middle latent space of search I don't find useful.
Glean is not a real business and is more of a VC back dealing play to buy revenue and fund some acquisition. Lucky for those who got in on it I guess early. Insane multiples for a not that interesting of a product or certainly one that is easily replicated, and the usefulness (internal docs) is one where businesses would cut costs easily as its not a profit center. Everything they offer is being commoditized.
I think the difference is that at Amazon service teams owned their service and code. At Google you had 10ks of engineers who never produced really any client impacting or customer facing code, so there was no need for blast radius protection or management, nor stringent testing. At Amazon even internal teams were often serving 100k end users and had expectations.
Too few doctors thing is problem with different bottlenecks... Mainly being that the US Residency program is funded by the gov't, and is competitive to get into. Many of the lesser schools and especially DO or carribean schools have graduates who are not able to get into residency on first try because there are not enough slots.
classic fallacy seen on engineering blogs all the time "we cut down {speed, cost} by writing something in {go, rust} without realizing that rewriting is also refactoring from the most explicit set of requirements. None of the tech debt or technical decisions are factored in.
you could just make a free git repo with this. Matter of fact some cloud providers even provide a lot of this setup. IMO the most difficult part for building is not this configurational infra, but integrating it into a well greased CI/CD deployment. Service meshes are kind of nothing and super simple to setup.
the only development (and it wasnt invented by neuralink) is flexible and smaller electrodes seem to take longer for scarring to take over than larger and more rigid ones. Right now, it seems that their only path is to help people with severe disabilities get better quality of life for a few years, which many people might take.
Accenture also performs weird menial tasks under guise of efficiency for large corps apparently unable to do it themselves. I would venture they will also roll out platform to enable corps to run their banal trainings on it (compliance etc)
I feel that the more immediate and impactful opportunity that people are doing is instead of scraping to get/understand content. LLM agents can just interactively navigate websites and perform actions. Parsing/Scraping can be brittle with changes, but an LLM agent to perform an action can just follow steps to search, click on results, and navigate like a human would
I would think it is intentional and brand strategy. OpenAI is such a force majeure that people will not know how to switch off of it if needed, makes their solutions more sticky. Other companies will probably adjust to their terminology just to keep up and make it easier for others to onboard.
Not really about the framing of these policies. What it really should be is 'new rules to limit exploitation of workers reduces exploitation'. What the US has too much of, is providing people just enough support that you have legions of people lining up for jobs that are effectively subsidized by tax payers. These services should not really exist at the current price point and scale.
I don't think those are really comparable. Spending a ton of money on a home theater is a large and static purchase. You put it in your home, and that's it, you can't move it easily or adapt it much. The Vision Pro, on top of having other functionalities, lets you have that home theater experience anywhere. Want o watch in bed? on the couch? outside on porch? waiting at the airport? People spend a lot more on phones because of the portability aspect.
They also vastly scaled back the number of services operating though. This whole thing of saying its stable with x << 100 % of staff is kind of nonsense when it its users, revenue, and features are also << what it used to be.
I would think you could improve your embedding space to address that issue, partially. Similarity search (as a result of some contrastive loss) definitely suffers at the tails and the OOD is pretty bad. That being said, you're more likely to have higher recall than a more classical technique.
no, this is just instructions coming down from the tippy top to do something. The reality is that politicians do not want to ever risk a cataclysmic incident happening under their guard with a disenfranchised population, I guess especially in more conscionable regions where voters won't react. The same thing happens even in DC a lot (high aesthetics budget to keep it clean) - The crux is that they know it will go back to the way it was so if you're not going to solve it doesn't seem like a good use of resources, although I'm sure residents might disagree. The only other thing that can propel it is an incident such as what happened in Seattle shooting and such then they go in as well.
While neat, this does not align with what I consider readable code. Outside of an IDE, this code can be quite hard to follow when you start to introduce tricks like these, especially since it's not paradigm in other languages.
not very useful, but it's getting better with XLA. It does not just "work" by default for any given model, and if you're missing an HLO that has an optimized kernel written you lose benefits of accelerated computing thus people won't use it. I would not put any effort into it until they truly have it where you just cast `to_device(..)`.
Managed services cannot be beat for moving quickly on infra that is not a part of your core competency. Because winding down is as simple as shutting down the instances.
But the point about saving not even right. It costs a lot in staffing to set up some of this infra, and if you're a serious company (revenue and expectant customers), you're going to need redundancy on these things anyway, so its more than 1 person (they can be responsible for multiple things)
Resources that are shared across taxpayers... They have to pay to heat and power their homes. They are also paying property taxes that support local schools, but they are not sending their kids to school, therefore benefitting everyone else with the tax dollars. They add a lot to tax base without adding to the cost. Even things such as sewage treatment and the roads of which they contribute to they use less of.
truly dystopian to live in a VC backed neighborhood where you don't own anything... Maybe if you were recent graduate for a few years. The whole idea of not needing a car is also not even that exciting. You don't exist in a vacuum in your tiny neighborhood in a larger city, you still probably need to get to work, US does not build good enough public transportation especially in these less dense cities. If you want to get out anywhere else in city or leave you'd probably want or need a car. Most people who live in NYC live close to people and surprise also do not need cars.
I actually really like what you are working on, it's a hugely unoptimized and addressed problem. If you could truly unlock potential it would be huge enabler for a large segment of businesses.
Almost all major cloud players and a few startups are spinning up massive GPU clusters. Saying that building compute creates almost intrinsic value is the greatest bridge selling I've seen in a while.
ah "multi cloud" yes we use Microsoft office for productivity and run our services on AWS.
Very few heavy applications run on multi-cloud, serving static and deployments at edge maybe to get geographic distribution. But this stat is mostly ridiculous.