ein0p

9 ms·

61.
by ein0p 1y ago
66.8% error rate reduction for o4-mini on AIME2025, and 21% error rate reduction on MMMU isn't "slightly higher". It'll be quite noticeable in practice.
62.
by ein0p 1y ago
These small models only make sense with "thinking" enabled. And once you enable that, much of the cost advantage vanishes, for output tokens.
63.
by ein0p 1y ago
How will it lock in the enterprise if its market share of enterprise customers is half that of Azure (Azure also sells OpenAI inference, btw), and one third that of AWS?
64.
by ein0p 1y ago
Absolutely decimated on metrics by o4-mini, straight out of the gate, and not even that much cheaper on output tokens (o4-mini's thinking can't be turned off IIRC).
65.
by ein0p 1y ago
What you need to do is unsubscribe entirely from all social media and stop reading news first thing in the morning. You have to remember that "Tobi" can't tell a pointer from a hole in the ground, and neither can Sam Altman.
66.
by ein0p 1y ago
Some guy who self-admittedly wrote no code for the last 11 years went to another startup, and we're supposed to hyperventilate? Is there any other way to "step down" than "abruptly"?
67.
by ein0p 1y ago
So good, they've made it illegal. Jokes aside though, aside from installing the Great Firewall of USA, I don't see how this could possibly work. And even then the cat is out of the bag. Models and code are speech, and there's
68.
by ein0p 1y ago
Take a look at their own paper or at many attempts to train something large with this. There's no replacement for displacement. If this actually worked without quality degradation literally everyone would be using this.
69.
by ein0p 1y ago
Um, yeah, rolling out something that hasn't been done before (large scale robotaxi service) is going to be slow, choppy, and hard. What did they think the analysis would show? Ask Google about how hard this is even with a handful of ta
70.
by ein0p 1y ago
This is over a year old. The sky did not come down, everyone didn't switch to this in spite of the "advantages". If you look into why, you'll see that it does, in fact, affect the metrics, and some more than others, and
71.
by ein0p 1y ago
On the one hand he's right, Apple needs a Musk of sorts, Jobs having kicked the bucket long ago. On the other, he has an axe to grind, and that axe is well known, and he grinds it at every opportunity. We get it, DHH, you don't li
72.
by ein0p 1y ago
Could be that they're negotiating with both in an attempt to create better price leverage. That's what I would do.
73.
by ein0p 1y ago
People sort of expect it to 100% subscribe to the right wing dogma now, but Elon apparently wasn't joking about it being "truth seeking". It seems pretty impartial to me, on some topics even "woke".
74.
by ein0p 1y ago
The closest Elon ever came to anything Hague-worthy is allowing Starlink to be used in Ukrainian attacks on Russian civilian infrastructure. I don't think the Hague would be interested in anything like that. And if his life is worthles
75.
by ein0p 1y ago
Does anyone have something like this for Xiaomi Band? I have one but the stock faces are an assault on my aesthetic sensibilities. Edit: apparently there is an app called EasyFace, but it doesn't seem to be open source, and it says it
76.
by ein0p 1y ago
You also can get Grok to fact check bullshit by tagging @grok and asking it a question about a post. Unfortunately this is not realtime as it can sometimes take up to an hour to respond, but I've found it to be pretty level headed in i
77.
by ein0p 1y ago
Good. I used to be a daily NPR listener, but it lost the plot years ago. Now every time I accidentally switch to it it's either 5 minutes of Trump hate or some "current thing" I don't care about.
78.
by ein0p 1y ago
That whole acquisition always was a head-scratcher to me. OK, you pay nearly 17B for the thing, and then do absolutely nothing with it. How does that work? Even during its best years 17B was big money for Intel.
79.
by ein0p 1y ago
No such thing as FU money. As you get more you want more, or if you don't then your spouse does. There is such a thing as "I can tolerate a 3-6 month job search no problem and not jump at the first opportunity" money. It make
80.
by ein0p 1y ago
The point is, for this to work, your system has to be connected to the internet. And if that's the case, there's no material difference between hosting it on prem and hosting it on GCP (where Google can promise you that it won&#x2
81.
by ein0p 1y ago
I don't have this, but I have noticed that I have trouble recalling names especially if I neglect my sleep for too long. It's not just names of persons, programming knowledge is affected too - I forget the names of functions, conf
82.
by ein0p 1y ago
That is beside the point. No lawyer will allow any "learning" on private data. Even without legal counsel, companies would be stupid to do it without compensation.
83.
by ein0p 1y ago
OP spent several years at Google. Kids or no kids, if they managed their finances well, they have a lot of latitude wrt next moves.
84.
by ein0p 1y ago
This is the way. Options aren't really worth much for the rank and file startup employees after about 7-10 hires. That fraction of a percent is just not going to be life changing unless it's the next OpenAI or something. For very
85.
by ein0p 1y ago
What a sudden change of heart. Thank you, DeepSeek!
86.
by ein0p 1y ago
The whole point of deploying such things on-prem is air-gapping it from Google and its "learning".
87.
by ein0p 1y ago
For this to take off it has to be supported by Nvidia, at feature parity with C++. Otherwise it'll remain a niche tool, unfortunately. I do hope Nvidia either supports this or rolls out their own alternative. I've been looking to
88.
by ein0p 1y ago
As an ex-Googler I say: blessing in disguise. When working at a $MEGACORP it's easy to think there's barren wasteland out there beyond the walls, so it's scary. But that is very much not so. I get that opportunities to work o
89.
by ein0p 1y ago
More degrees of freedom, easier collision avoidance. The mechanical part is a lot harder, of course, but the "self driving" part is a lot easier. We had autopilot and auto-land in planes well before we had anything of the sort in
90.
by ein0p 1y ago
I hear from my OpenAI contacts that the next wave of thinking models are going to blow the socks off Google. In some ways they already do (speech, images). So this lead will likely be short lived. That said, in the meantime I did get paid A

More ›