Nice webpage/design, but what is the advantage over Openrouter that already has ~trillions tokens/month in traffic?
1,414 karma · joined June 13, 2021
What difference are you seeing from these models that makes it better?
I say this as someone considering shelling out $200 for ChatGPT pro for this.
Assume 1,000 customers each generating $2m in ARR with contracts. That’s $2 billion. Assume generous 6x ARR valuation, that’s $12 billion.
Where is this $20 billion premium coming from? How could the board approve this? How is this fair to shareholders?
Heck, as a minor shareholder in GOOG, I don’t find this financially responsible at all.
I can’t help but think sometimes these tech acquisitions have some hint of nepotism/deeper underlying motivations behind them than meets the eye.
- Actually, blatantly faking videos of Tesla FSD.
- Elon claiming unsupervised NY to LA would be possible by end of 2016 (almost 10 years ago).
- Elon claiming all cars sold today (years ago) have required hardware to become robotaxis.
- Tesla salesmen of the past telling customers their cars were investments because they would become robotaxis that would make money.
Theranos lasted 15 years, so I'm surprised $TSLA is being called this early.
As someone who personally knows people in the industry, I'm pretty confident in saying that TSLA will never actually deliver any type of robust, safe autonomous driving system. The organization itself is just not capable of it anymore.
All that are left are people trying to game metrics and not get fired, that type of culture just can't produce it.
Am I working hard at Meta? You bet I am.
For median SWE salary and 0.001% in “options” that will likely never have a liquidity opportunity? I’m not so sure…
Things like: counters, news feeds, chat messages, etc
The cost of delivery for doing these things well with a LSM based DB or RDB might actually be higher than Redis. Meaning: you would need more CPUs/memory to deliver this functionality, at scale, than you would with Redis, because of all the overhead of the underlying DB engine.
But for 99% of places that aren’t FAANG, that is fine actually. Anything under like 10k QPS and you can do it in MySQL in the dumbest way possible and no one would ever notice.
Meanwhile property prices and 401ks are up, old people (the largest cohort of voters) really just don't care.
I wish people had more empathy for others, but they don't. They mostly only care about taxes and their personal wealth.
This level of personal greed-individualism is partly what makes America great in the good times, but awful when things get bad.
Oh well, as long as he's "one of us" right? Who cares about national security, privacy, etc... Trump will pardon him and DOGE will have much win much budget cuts amirite?
if depth > 5 and if sem_hash(content) in hist: return
What are people writing that requires such fancy/extensive usage of the borrow checker?
I can't even remember the last time I had to use a shared_ptr... unique_ptr and other general RAII practices have been more than enough for our codebase.
> Low paying government job is not secure.
Spends like 6 minutes with a creepy cardboard face of MrBeast in the clinic and performing "eye surgery" on him tf?
It's still not clear to me what happened; did Beast interact with him and tell him he would pay and never did? Did Beast just show up to his clinic and make the whole thing up?
This only has 17k views and was uploaded 8 months ago. It's because it's incoherent, the story or situation is not clear at all. With effective storytelling this could have gone viral, ironically, effective storytelling like the kind that MrBeast used.
Taking raw embeddings and then storing them into vector databases, would be like if you took raw n-grams of your text and put them into a database for search.
Storing documents makes much more sense.
Supervised learning you train on pairs of (x, y) where x is your input (title/post text/metadata) and y is the output score.
Naively, it's a linear regression model, Y = b0 + b1x1 + b2x2 + b3x3. Where b0 is your bias ("a floor for score points"), and b1, b2, and b3 are bias terms for the actual data of the post. You can solve this, closed form, and find the b1/b2/b3 that minimize the error of fitting to Y.
How do these equations change with RL? I always assumed RL was a multi-step process where actions are taken to get to a reward. If there is only 1 step/decision, to produce a "random" score, it feels much like supervised learning.
In English for example, learning the 800 most common words, you can understand 75% of the language: https://www.bbc.com/news/world-44569277.
I'd love to start fresh on a new language, take 800 new words, try to learn 10 a day, and see where I get after 3 months. Can I really understand 75% of text if I have perfect recall of those 800 words?
The vast majority of human experience has been tough. Arguably the vast experience of most life is pretty cruel.
Should they just donate the wealth of their country to some other one? Would you like to donate some of your personal income to the nice homeless people outside?
The leaps from people build dumb glass city => world is bad => people building dumb glass city are bad because world is bad aren’t obvious to me.
It's easy to forget that it's still such a new and early field. The modern version of CS started in what... the 50s, earliest? There are many people alive today who are older than the study of CS itself.
You'd be surprised at how difficult it is to get liquidity for that stuff, often there are limits to the amount you can liquidate, some can't sell on private marketplace, some can only sell every once in a blue moon event, etc. This is all without mentioning that the "valuation" itself is typically pretty speculative.
Levels.fyi is treating this equity the same as public company RSUs, which is not the same at all.
Basically you pay them a lot of money just so they puff you up on the phone and make you feel important.
Someone will come in and argue, "but they can help mitigate risk, and and alignment with financial goals and timelines." ... Yeah, and so does just investing in index funds yourself and reducing your "risk" by allocating some % into bonds.
Other examples I can think of (from my Gen Z experience, growing up in SoCal):
- Elementary school: Making custom action replay codes, hacking game saves with programs, CheatEngine/memory/hex editor and following YouTube tutorials, Javascript "document.contentEditable=true" hack and changing stuff on websites, pressing F12 and changing random javascript code until something interesting happens or breaks.
- Middle school: making sites on Weebly/freewebs, embedding chats and flash games on them, sharing them during computer class
- High school: Making PHP sites/vbulliten/Newgrounds/flash games, later iOS apps
I wasn't the only one doing these things. There were always like 3 other kids like me in any classroom that would do the same things.
Most of us ended up becoming passionate SWEs, besides one that became an accountant.
Did they maybe expand to a tree during training to learn more robust reasoning? Maybe. But it still comes down to a regular transformer model at inference time.
1) The "bitter lesson" may not be true, and there is a fundamental limit to transformer intelligence.
2) The "bitter lesson" is true, and there just isn't enough data/compute/energy to train AGI.
All the cognition should be happening inside the transformer. Attention is all you need. The possible cognition and reasoning occurring "inside" in high dimensions is much more advanced than any possible cognition that you output into text tokens.
This feels like a sidequest/hack on what was otherwise a promising path to AGI.
Probably 1% of users will get access to it, with a 20/message a day rate limit. Until early next year.
Idk if I'm "feeling the AGI" if I'm being honest.
Also... telling that they choose to benchmark against CodeForces rather than SWE-bench.
100M context window means it can probably store everything you’ve ever told it for years.
Couple this with multimodal capabilities, like a robot encoding vision and audio into tokens, you can get autonomous assistants than learn your house/habits/chores really quickly.
I'd happily pay 10x this ($3,000) to never do laundry again if it meant nicely folded clothes would always await me in my closet.