> Algorithms to live by - Brian Christian [and Tom Griffiths]
I think it would have been very difficult for Brian Christian alone to produce Algorithms to Live by: Tom Griffiths must have had a very strong presence in it.
I didn't graduate high school and I can program in multiple programming languages, and picking up a new one is pretty easy. I had to convert "math functions" into code before I could understand what the symbols meant when I was learning how to write ML software.
Imo you learn the academic lingo, and it can make it easier to communicate with other programmers, but you can just pick it up. School is for chumps.
Wikipedia is trash, always has been. A small group of fanatics run that site and only allow information they deem acceptable. Using their information has always been a bad idea.
Haha "Anonymous" to your neighbor maybe, if they don't work for 5Is or a 3 letter agency, or Google, or Facebook, or the ccp or ... tor is compromised, you didn't Not get caught, you're not a target yet.
You're the second person in this thread to claim that Tor is compromised, without giving any explanation. This isn't productive discussion. Care to elaborate in what exact manner is Tor compromised? Bonus points for not hitting anything from the known list of speculative attacks. [1]
Traffic correlation: All the large state actors are well capable of recording every single IP transaction between devices. You can create detailed correlation maps from these transactions. Considering that this wouldn't cost much for state actors to implement, one has to assume such traffic correlation systems currently exists.
Node compromise: It costs less than 5 dollars a month to create a TOR node. There are currently ~8000 TOR nodes/relays in existence. That is 40k USD per month (at most). Do you really believe state actors can't afford 40k USD per month to compromise the vast majority of TOR nodes? Even a single millionaire can compromise the vast majority of TOR nodes.
Another problem is that TOR is an outdated privacy tech. considering modern state actor capabilities. Mixer networks + network jitter is necessary to protect privacy at this stage, yet no such project exists yet.
TOR is not a good option for privacy. Currently only valid option for privacy is external Wi-Fi jacking and ensuring you don't send any private info like CPUID.
Or alternatively, you can hack routers/computers and put your own TOR nodes in them, then you can only use these known nodes.
> Note: even though it originally came from an acronym, Tor is not spelled "TOR". Only the first letter is capitalized. In fact, we can usually spot people who haven't read any of our website (and have instead learned everything they know about Tor from news articles) by the fact that they spell it wrong.
We all know which one you fall under.
> Node compromise: It costs less than 5 dollars a month to create a TOR node. There are currently ~8000 TOR nodes/relays in existence. That is 40k USD per month (at most). Do you really believe state actors can't afford 40k USD per month to compromise the vast majority of TOR nodes? Even a single millionaire can compromise the vast majority of TOR nodes.
As someone who has ran a Tor exit node for 5+ years, lolol at this statement. Pure ignorance.
Common sense is not the same as actual understanding. Tor is not limited to exit nodes, the most valuable thing is the onion network with millions of users. Both are being actively monitored for anomalies by the devs, with multiple thwarted attempts to subvert it in the past.
> TOR is not a good option for privacy. Currently only valid option for privacy is external Wi-Fi jacking and ensuring you don't send any private info like CPUID.
What is known for sure is that high profile drug dealers are using it without being caught for years. All known cases are related to either poor OPSEC, client/server 0-days, classic real-life investigations, or known attacks Tor can't protect from (correlation of the large amounts of onion server traffic, for example, which is not that easy as you make it sound). It is entirely possible that somebody was caught using unknown or unavoidable attacks, but no such case is known for sure at this time.
> Mixer networks + network jitter is necessary to protect privacy at this stage, yet no such project exists yet.
Tor does use packet shuffling and delays to protect from timing attacks to an extent. It's less advanced than I2P which also mixes the traffic, but has a much larger client pool and a unified browser used by nearly everyone, which provides users with huge buckets to blend into. There are also several delayed onion message services available.
Of course it is susceptible to certain kinds of attacks you have to be aware about. This isn't the same as "Tor doesn't provide anonymity" or "Tor is compromised".
People who have used Tor who have been deanonymized in every single case that we are aware of made other mistakes ("altoid", "pimp_alex_91@hotmail.com", etc) or the vulnerability was in an outdated version of Tor Browser (basically Firefox). Of course you can say there was parallel construction, maybe there was in some cases, but if they didn't make those other mistakes there would be no case that could be brought against them because no one is going to be convicted on the basis of some probabilistic attack against Tor that the government doesn't want to reveal to begin with.
These are just 0-days in the Tor Browser/Firefox, not Tor. While in practice Tor Browser cannot be decoupled from the network (it provides users with large enough crowd to blend into), it can be surrounded with additional security layers. If you're a high-value target worthy of wasting a 0-day on you, running Tor Browser on a general-purpose system might be a bad idea; there are distros like Whonix specifically tailored for secure communication - they are far less vulnerable to this kind of attack.
I've been training large 65b models on "rent for N hours" systems for less than 1k per customized model. Then fine tuning those to be whatever I want for even cheaper.
2 months since gpt 4.
This ride has only just started, fasten your whatevers.
Finetuning cost are nowhere near representative of the cost to pre-train those models.
Trying to replicate the quality of GPT-3 from scratch, using all the tricks and training optimizations in the books that are available now but weren't used during GPT-3 actual training, will still cost you north of $500K, and that's being extremly optimistic.
GPT-4 level model would be at least 10x this using the same optimism (meaning you are managing to train it for much cheaper than OpenAI).
And That's just pure hardware cost, the team you need to actually makes this happen is going to be very expensive as well.
edit: To quantify how "extremely optimistic" that is, the very model you are finetuning, which I assume is Llama 65B, would cost around ~$18M to train on google cloud assuming you get a 50% discount on their listed GPU prices (2048 A100 GPUs for 5 months). And that's not even GPT-4 level.
As I stated in my comment, $5M is assuming you can do a much much better job than OpenAI at optimizing your training, only need to make a single training run, your employees salaries are $0, and you get a clean dataset for essentially free.
Real cost is 10-20x that.
That's still a good investment though. But the issue is you could very well sink $50M into this endeavour and end up with a model that actually is not really good and gets rendered useless by an open-source model that gets released 1 month later.
OpenAI truly has unique expertise in this field that is very, very hard to replicate.
No I'm not, it's the full model on 8 gpus for a couple hundred.
After training I fine tune for chats but mostly command and control tools, and then you fine-tune for application.
reply