432 karma · joined January 16, 2017
What’s the angle for stripe , electrify over tokens exchange is the new money flow , and stripe wants to monetize it. 5% tax on any llm token is an amazing deal
Spacex even have multi model one
If you willing to share to no zdr, meta is waaaaaay cheaper vs Kimi.
With recent offerings from spacex and meta , I hardly imagine why would you pay money to any Chinese vendor it’s not as cheap and it’s not as intelligent neither .
Maybe deepseek is an exception , but it’s only good for narrow use cases that probably goes into modal.com and other gpu + fine tune me easy vendors , not vanilla dumb but cheap model .
I used auto in cursor it’s much faster va Claude code and as good.
Image is content , ai can generate art level images and an obvious slop . To read or or not is absolutely up to reader , not someone’s entitled opinion .
Spacex > I am talking about their revenue , and they did resell their gpus with margin when their model clearly underperformed .
Stocks … two weeks ago it was amazing now it’s slightly under water , who knows what that be in 6 months. It does not matter for this conversation.
The only real problem to them they killing the market they have most money off. Software . But ai already in a good spot to displace Microsoft office after that other industries , finance , lawyers , medicine etc
Chinese models pushes prices down and quality up, that makes GPU-based automation more affordable, while covering more and more cases to automate.
You can debate that llm producers will go bankrupt, some of them at least for sure.
How do you lose in this market if you do gpu?
In fact spacex story tells you next : investments in infrastructure is the best investment. If OpenAI or anthropic have committed infra in the worst case scenario they can re-sell it with margin .
The only way it will not payout suddenly we wake up in the world where ai fails to deliver . Which does not seems to be the case .
If you have properties of the market where your costs will go down , the size of the market will increase and you are top contender. How is that a bubble or a bad market ?
Sure you have risks of underperforming and lose the competition, but how is that different from any business in the world ?
Users go on vacation, they slack off, they spend the day talking to each other. There are very few people who are really effective at burning tokens. how do you know the ratio? do you have insides? No :)
The biggest target is enterprise, and the economics for an LLM vendor look like this: price per token = R&D + inference + infra investments. When you buy a subscription, you are quite often buying a year ahead. That lets the vendor predict future infra investments against hard commitments, and sell expensive per token pricing to everyone else. And when a hard commitment sits unused because the user is busy, they sell it twice. It is loyalty in exchange for predictability, in exchange for the promise to always deliver SOTA to users.
Vendors control the harness. Tomorrow they simply roll out a router where reading the code and doing the final edits goes to a cheaper model, and their math suddenly becomes very sexy.
Isn't that hard to predict that their economic model is very easy to tune? and this is just first baby steps.
I personally pay per token ( do not have subs for work ). I did have once a $25k/mo worth of tokens, since i knew it was free so i was doing crazy experiments. Now , 2 month later, my bill was barely $1.5k since i moved into different stage with project. I do have team members who burn $500-600. pre router, pre optimization.
I switched recently to grok 4.5 and cursor router and my bill will go even further down. It rotates 4-5 different vendor models cheap and expensive too, depends on the task. Routers will flip entire LLM economy upside down.
I think building it , will be better. Custom solution can have any memory you dream of integrations, it can recommend skills with some nesting or not the call can combine all instructions or not if you want or structure it in custom objects.
Also if you go custom it can be part of your release process. Usually i'm pro libraries, but in this case hard to imagine going with such tool. Connecting it properly to self learning loops might be a challenge too.
- almost every book try to stretch core idea into book size format
- unique ideas are rare people attack them under different angels
- a later phenomena : their believes almost predict entire book, outcomes etc, brainwash impact is real
so not sure, how ai slop is better vs book slop, at least with ai you can distill the idea, with the book, you have to spend 10-40 hours to digest average, absolutely non fresh ideas, that author brought in just to sell that book, otherwise it would be magazine article worth.
Q: Did you just way over-optimize for a specific CPU and tokenizer? How is it so fast? No, I way over-optimized for every combination of these! The results are very consistent across CPUs (modern x86 and ARM), and across specific tokenizers.
The major improvements are in optimizing heavily an implementation that usually is outsourced to a Regex engine (pretokenization) using SIMD, minimizing branching and other tricks, as well as heavily optimizing caching of pretoken mappings (if a word has been seen before, look it up its encoded tokens efficiently). Caching is a very hard problem in this domain since the cache grows very quickly, and pretoken distributions are very long-tailed.
Finally, interactions with Python are minimized, and threads have minimal interactions with each other.
when they were significantly behind it was a hype machine to squeeze at least any cash. GLM CEO openly said, that open source is a hype engine for them.
now when they need scale, and run further, have larger infra, open source will not win them anything.
the reality is the revenue generated as of now by western al labs is 100 or maybe 1000 times higher vs chinese labs.
As a business, open source a model is a desperate move. It's a 0 benefit except getting recognition. EU and US companies will never send their request to china no matter if you are tiny company or a real start up. You always deal with someone sensitive that will block you doing so. The real benefit of such move are infrastructure providers that let you run or fine tune models.
Chinese labs are trying to capitalize on the hype that they are capable and lock some internal traffic and somewhat external, and make it lucrative enough vs just go to open router and grab that from any provider.
you can pin your Bluetooth app with drag-n drop , never used car link so not sure if that's an app or an icon.
glovebox is the real annoying item i can confirm is a 100% weird design choice. I still love pin in valet mode so that my items are not stolen.
Why on earth everytime i sit in the car i have to neurotically plug the cable. I had situations i forgot and in the middle of the ride i started to look for cable or try to wire car play, simply because i figure out i don't know where to go .
This type of crap is not convenient and simply dangerous.
what proper car should offer: keyless enter; keyless profile adjustment including seats, heat, music profiles, etc wheel etc; keyless start and go;
one of the best tiny features i have. In the morning i drive to work; i can go several ways there depends on traffic, My car knows that in the morning i go to work; GPS is on, traffic is calculated.
I open the door , my music started to play, I press gas and drive; zero cables, zero pairing ; open the door and drive.
Car is usable, i use it to open trunk frunk and check the status the door/trunk. I'll tell you more , that type of visualisation exists in almost every single modern car.
their navigation is just the best not sure what problem are you talking, it's FSD knows better where to turn vs myself in unfamiliar area. I do have less miss exists on fsd with myself. Yesterday i 0 touch 1h 30 min drive from long island all the way to a very busy manhattan street. tons of exits, complicated connections etc.
It's not even comparable. Everything else feels like a horse carriage vs space ship.
Joe mode ( silent your car when baby sleeps)
automatic profiles that are connected all the way to what music played and in what device, headphones seamlessly stop playing after i go into the car from the gym
send navigation to car from any device , while I'm driving my gf or friend can send it, so i do not distract myself
proper navigation , order management etc as part of the vertical integration, including re-routing to less busy charger etc.
- when you need to re-pair Bluetooth
- when you forget the cable to charge and you need to drive
- when you want to share your car to someone and they need to spend 5 minutes to accept every single ToS possible to simply put a GPS
- several people with phones paired before, now you dealing with complete random
you name it.
- you listen music and you need to go out to buy something while others in the car
None of these problems exist if you have a decent, dedicated computer in the car that just works, it knows profiles, it does need you to be always on wire, or on the line.