931–940 of 965 posts
Earlier quoted context omitted.
This is so frustrating. Why don't they just make it available everywhere?
Because the EU requires them not to: https://ec.europa.eu/information_society/newsroom/image/docu...
Public benchmarks are broadly indicative, but devs really should run custom benchmarks on their own use cases. Replicate created a Llama 3 API [0] very quickly. This can be used to run simple benchmarks with promptfoo [1] comparing Llama 3 vs Mixtral, GPT, Claude, and others: prompts: - 'Answer this programming question concisely: {{ask}}' providers: - replicate:meta/meta-llama-3-8b-instruct - replicate:meta/meta-lla…
We had some issues with the problems with the vocab (showing "assistant" at the end of responses), but it should be working now. ollama run llama3 We're pushing the various quantizations and the text/70b models.
Earlier quoted context omitted.
In today's Dwarkesh interview, Zuckerberg talks about energy becoming a limit for future models before cost or access to hardware does. Apparently current largest datacenters consume about 100MW, but Zuck is considering future ones consuming 1GW which is the output of typical nuclear reactor! So, yeah, unless you own your own world-class datacenter, complete with the nuclear reactor necessary to power the training ru…
A bit odd to see this downvoted... I'm not exactly a HN newbie, but still haven't fully grasped the reasons people often downvote here - simply not liking something (regardless of relevance or correctness) seems to often be the case, and perhaps sometimes even more petty reasons. I think Zuck's discussion of energy being the limiting factor was one of the more interesting and surprising things to come out of the Dwar…
Earlier quoted context omitted.
Do they really need “free RLHF”? As I understand it, RLHF needs relatively little data to work and its quality matters - I would expect paid and trained labellers to do a much better job than Joey Keyboard clicking past a “which helped you more” prompt whilst trying to generate an email.
Absolutely. Modern captchas are self driving object labelers; you just need a few to "agree" to know what the right answer is.
They've got a console for it as well, https://www.meta.ai/ And announcing a lot of integration across the Meta product suite, https://about.fb.com/news/2024/04/meta-ai-assistant-built-wi... Neglected to include comparisons against GPT-4-Turbo or Claude Opus, so I guess it's far from being a frontier model. We'll see how it fares in the LLM Arena.
Also added Llama 3 70B to our coding copilot https://www.double.bot if anyone wants to try it for coding within their IDE and not just chat in the console
Earlier quoted context omitted.
We had some issues with the problems with the vocab (showing "assistant" at the end of responses), but it should be working now. ollama run llama3 We're pushing the various quantizations and the text/70b models.
What's the reason behind "assistant" showing up?
Earlier quoted context omitted.
There's a difference to being a good chatshow/podcast host and a journalist holding someone's feet to the fire! Dwarkesh is excellent at what he does - lots of research beforehand (which is how he lands these great guests), but then lets the guest do most of the talking, and encourages them to expand on what they are saying. It you are critisizing the guest or giving them too much push back, then they are going to cl…
I decided to listen to a Dwarkesh episode as a result of this thread. I chose the Eliezer Yudkowsky episode. After 90 minutes, Dwarkesh is raising one of the same 3 objections for the n-teenth time, instead of leading the conversation in an interesting direction. If his other AI episodes are in the vein as other comments describe, then this does seem to be plain old positive AGI optimism bias rather than some special…
The main strength of Dwarkesh is the caliber of guests he is able to attract, especially for being so new to the game. Apparently he'll research a potential guest for a couple of weeks before cold e-mailing them with some of his researched questions and asking if they'll come on his podcast, and gets a very high acceptance rate since the guests appreciate the questions and effort he has put into it (e.g. maybe Zuck enjoying being asked about Augustus, and not just about some typical FaceBook fare).
If you were inclined to give him another try, then I'd recommend the Richard Rhodes or Dario Amodei episodes, not because of any great Dwarkesh interviewing skills, but because of what the guests have to say. If you are a techie then the Sholto + Bricken one is also good - for same reason.
As far as AI optimism, I gather Dwarkesh has moved to SF, so that maybe goes with the territory (and some of his friends - like Sholto + Bricken - being in the AGI field). While arguably being a bit too deferential, he did at least give some pushback to Zuck on AI safety issues such as Meta's apparent lack of any "safe scaling" tests, and questioning how Zucks "increased AI safety via democratization" applied to bio threats (how is putting capability to build bio weapons in hands of a bad actor mitigated by others having AI too).
Earlier quoted context omitted.
The world at large seems to hate Zuck but it’s good to hear from people familiar with software engineering and who understand just how significant his contributions to open source and raising salaries have been through Facebook and now Meta.
A person (or a company) can be two very different things at the same time. It's undeniable as you say that there have been a lot of high-profile open source innovations coming from Facebook (ReactJS, LLaMA, HHVM, ...), but the price that society at large paid for all of this is not insignificant either, and Meta hasn't meaningfully apologized for the worst of it.
Basically I don’t think their privacy nightmare stuff detracts from what the brain trust of engineers over there have been doing in the open source world.