sigmoid10

8 ms·

121.
by sigmoid10 3mo ago
But it does cover the state described in the top comment. A<-B is never going to be easily retrieved if you only experienced A->B during training, regardless if you are a human neural network or an artificial one. Also, you need to de
122.
by sigmoid10 3mo ago
It certainly holds true for humans. The brain stores relational information in a sequential pattern that is not automatically reversible. One of the best examples is the alphabet. Everyone learns it in school, so the pattern A->B->C-&
123.
by sigmoid10 3mo ago
If you look at public benchmarks like ExploitBench [1], then you'll see this is mostly a question of token budget. Once you give it sufficient tokens to burn, GPT 5.5 is roughly as good as Mythos when it comes to finding bugs and build
124.
by sigmoid10 3mo ago
Youtube channels are also getting hyper-monetized now. Private equity firms finally learned that some of these informational channels draw a huge crowd of loyal viewers with a very specific kind of technical interest and have built high lev
125.
by sigmoid10 3mo ago
Maybe project managers will finally get to experience flow states?
126.
by sigmoid10 3mo ago
This is not some fundamental problem. There are amazing options that don't suffer this downside. Most people are just too lazy and want something that works out of the box, so they accept or ignore the inherent risks of tools like Clau
127.
by sigmoid10 3mo ago
Also should note that this is a working paper and not peer reviewed. Working papers are circulated to invite comments and discussion. Until some experts confirm their methodology I wouldn't take any of this for granted.
128.
by sigmoid10 3mo ago
That was only true before AI overview. Now all that info they scraped from public websites and fed into their models will still be available without anyone ever leaving Google's services.
129.
by sigmoid10 3mo ago
Well, because of things like COPPA I imagine most companies wouldn't want to risk anything here. So unless you have a website that is somehow guaranteed 100% blocking all traffic from the US or American citizens, you may as well implem
130.
by sigmoid10 3mo ago
If you linked to the actual source of the study [1] instead of a random blog only talks about the result, you would see the big banner that the authors put there noting that the study is horribly outdated. Current models do make developers
131.
by sigmoid10 3mo ago
>It is suspicious to me that "age assurance" is trending EXACTLY as AI agents become capable of autonomously operating It is not, because your premise is false. This whole thing has been going on for as long as kids have been o
132.
by sigmoid10 3mo ago
Because of the introduction of AI overview, click-through rates are dropping like crazy. Up to 70% of Google searches now end without any clicks to third party websites. So most search users already stay completely within Google's ecos
133.
by sigmoid10 3mo ago
This case was about PriceRunner, a price comparison platform that was suffering from Google prioritizing its own platform in search results. Klarna just happens to be the owner.
134.
by sigmoid10 3mo ago
Valuations are not permanent. Amazon dropped 90% during the dotcom bubble. And there is always another financial crisis coming.
135.
by sigmoid10 3mo ago
I'm thinking more of EULAs. Even if Anthropic somehow wedges this into their TOS, it might still be illegal. For example, in many US states this could potentially be classified as consumer fraud. You can't just sell one thing and
136.
by sigmoid10 3mo ago
Probably because the upgrades to the collider are so significant that it will be called the HL-LHC afterwards.
137.
by sigmoid10 3mo ago
As I said above, if you are worried about privacy while hooking up Claude Code, you need to reevaluate your understanding of this technology.
138.
by sigmoid10 3mo ago
If they only collect the data for analysis I guess this is fine (they already get way more sensitive data from users anyways, so if privacy is your concern you've made the mistake many steps ago). The much more interesting question is
139.
by sigmoid10 3mo ago
Yikes. Good to know that labor shares used to rebound after crises, but since the 2000s and the dotcom bubble it has basically been downhill only. So don't expect any of this to get better unless we roll back technology to the last mil
140.
by sigmoid10 3mo ago
That is not a problem for LLMs, because in practice floating point inaccuracies (in particular after exponentiation) prevent values from being exactly equal. That's why greedy sampling generally produces deterministic output for LLMs.
141.
by sigmoid10 3mo ago
The point is that the case T=0 doesn't just "exist" as a special code branch - it is still well defined mathematically without any change to the output function. What the above comment refers to with the extra "if"
142.
by sigmoid10 3mo ago
>in theory theory, temperature 0 doesn't really exist. It does exist very much, even if you go to pure math. Look at the softmax function and take the limit as T->0. It becomes a dirac-delta function. I.e. in a discrete setting (
143.
by sigmoid10 4mo ago
OpenAI was already holding models back because "dAnGeR" before anyone knew or cared about them. It's always been a PR gag and Anthropic just so happens to be better at marketing than making frontier models available to a gene
144.
by sigmoid10 4mo ago
At least they plan to give the public all versions. Feels infinitely better than whatever the hell is happening at Anthropic. > "Yeah, we've got the absolute best model out there. Trust us. Truly scary." > "O-ok?
145.
by sigmoid10 4mo ago
And even if you did have light-years of super dense detector material lying around, the energy deposited by such a neutrino is so tiny that you wouldn't be able to register the hit.
146.
by sigmoid10 4mo ago
I wouldn't resort to language examples, as the other comments show how this gets imprecise and lost in semantic details quickly. Instead think of a basic logic example: Consider an OR gate with inputs A and B and output X. If B=1 that
147.
by sigmoid10 4mo ago
>Some amount of knowledge is required for reasoning. This is the root of problem. If you think about STEM universities, they don't really teach you things you need in the real world. They teach you what you need to know in order to
148.
by sigmoid10 4mo ago
I suppose it is both. Basically all frontier models are inference-time compute bound thanks to reasoning. And actual reasoning traces are locked behind closed doors at all American labs. So whenever they want to push a new model and need to
149.
by sigmoid10 4mo ago
You could also use the responses api which stores all message contents (including reasoning) on OAI servers. This has been possible for quite a while now. Encryption is only necessary if you really care about local storage (which is differe
150.
by sigmoid10 4mo ago
I guess they still use a tokenizer? Why would this kind of issue be solved? The model fundamentally can't see the word character by character like you do. For o200k tokenizers for example, what the model sees are 3 tokens: [302, 1618,

More ›