Any idea what "output token efficiency" refers to?
Gemini Flash is billed by number of input/output tokens, which I assume is fixed for the same output, so I'm struggling to understand how it could result in lower cost. Unless of course they have changed tokenization in the new version?
Were you able to identify the manufacturer and model/revision of the failing motherboards? This would be extremely helpful when shopping for seconds hand servers.
LangChain is considered complicated to get started with despite offering probably the widest amount of functionality. If you are already comfortable with LangChain you are free to ignore that.
Take the announcement with a grain of salt. From German physicist Sabine Hoffenfelder:
> The particular calculation in question is to produce a random distribution. The result of this calculation has no practical use.
>
> They use this particular problem because it has been formally proven (with some technical caveats) that the calculation is difficult to do on a conventional computer (because it uses a lot of entanglement). That also allows them to say things like "this would have taken a septillion years on a conventional computer" etc.
>
> It's exactly the same calculation that they did in 2019 on a ca 50 qubit chip. In case you didn't follow that, Google's 2019 quantum supremacy claim was questioned by IBM pretty much as soon as the claim was made and a few years later a group said they did it on a conventional computer in a similar time.
TBH you need to take the youtube influencer Sabine Hoffenfelder with a bigger grain of salt. She has converted to mainly posting clickbait youtube stuff over the last years (unfortunately, she was interesting to listen to earlier).
The RCS is a common benchmark with no practical value, as is stated several times in the blog announcement as well. It's used because if a quantum computer can't do that, it can't do any other calculation either.
The main contribution here seems to be what they indeed put first, which is the error correction scaling.
I think simplifying her to 'youtube influencer' is unfair - she is a doctor of theoretical physics with a specialism in quantum gravity who produces science content for youtube. She knows the field enough to comment.
She doesn't even say that this isn't a big leap (she says it's very impressive - just not the sort of leap that means that there are now practical applications for quantum computers, and that a pinch of salt is required on the claim of comparisons to a conventional computer due to the 2019 paper with a similar benchmark).
As a counterpoint, she recently reviewed a paper in one of her recent videos and completely tore it to shreds as apparently the math was full of absolute nonsense.
This was a fascinating watch, and not the kind of content that is easy to find. Besides videos like that one, I enjoy her videos as fun way to absorb critical takes on interesting science news.
Maybe she is controversial for being active and opinionated on social media, but we need more science influencers and educators like her, who don't just repeat the news without offering us context and interpretation.
I think I know which you mean, but TBH that paper read as an AI auto-generated troll paper that any physics undergrad should be able to dissect :) It was a bit fun to watch though and sure sometimes you need to provide some fun content as well!
How would she not survive on YouTube? Would they block her posts if she didn't use a misleading title and thumbnail (shout-out to dearrow for those who despise this practice)?
>Would they block her posts if she didn't use a misleading title and thumbnail
So, the way youtube works is that every single creator is in an adversarial competition for your attention and time. More content is uploaded than can be consumed (profitably, from Youtube's point of view). Every video you watch is a "victory" for that video's creator, and a loss for many others.
Every single time youtube shows you a screen full of thumbnails, it's running a race. Whichever video you pick will be shown to more users, while the videos you don't pick get punished and downranked in the algorithm. If a Youtube creator's video is shown to enough people without getting clicked on, ie has a low clickthrough rate, it literally stops being shown to people.
Youtube will even do this to channels you have explicitly subscribed to, which they barely use as a signal for recommendations nowadays.
Every single creator has said that clickbait thumbnails have better performance than otherwise. If other creators are using clickbait thumbnails, you will be at a natural disadvantage if you do not. There are not enough users who hate clickbait to drive any sort of signal to the algorithm(s).
If you as a creator have enough videos in a row that do not do well, you will find your entire channel basically stops getting recommended.
It's entirely a tragedy of the commons problem: If every user stopped simultaneously, nobody would suffer, but any defectors would benefit, so they won't stop simultaneously.
Youtube itself could trivially stop this, but in reality they love it, because they have absolutely run tests, and clickbait thumbnails drive more engagement than normal thumbnails. This is why they provide ample tooling to creators to A/B test thumbnails, help make better clickbait etc, and zero tooling around providing viewers a way to avoid clickbait thumbnails, which would be trivial to provide as an "alternative thumbnail" setting for creators and viewers.
Sabine is literally driving herself down an anti-science echochamber though. Maybe she can't see it, but it's very clear from the outside what is happening. She has literally said that "90% of the science that your tax dollars pay for is bullshit" which is absurd hyperbole, and something that a PHYSICIST cannot say about all fields full stop. It's literally https://xkcd.com/793/
This is in a world where people view youtube through the interface which is obviously most people. I have feeds that I subscribe to and some of them are must-see for me, so no matter what they use as their bait I am watching it. Fortunately there exists smarttube, dearrow and sponsorblock for people like me who just want to watch the stuff they've subscribed to and not whatever advertises the best.
> they provide ample tooling to creators to A/B test thumbnails
They do? For many years, I made my living from YouTube. This was always a feature that people wanted, but that didn’t exist. It’s been a year-plus since I’ve actively engaged on YouTube as a creator. Is this a recent change?
Just because she is a YouTuber doesn't diminish her other credentials, just as she is incetivised to do clickbait, so are actual scientific communication outlets such as nature, and the more clicky they are the more downloads and citation they will acquire. Incentives change content but don't directly detract from someone's expertise. See: the fact that most universities now publish some lectures on YouTube, it doesn't make the content any less true.
she was right the first time when they announced this in 2019 and this time even they admit in their own press release:
> Of course, as happened after we announced the first beyond-classical computation in 2019, we expect classical computers to keep improving on this benchmark
As IBM showed their estimate of classical computer time is taken out of their a**es.
They explicitly cover all of these caveats in the announcement.
Problems that benefit from quantum computing as far as I'm aware have their own formal language class, so it's also not like you have to consider Sabine's or any other person's thoughts and feelings on the subject - it is formally demonstrated that such problems exist.
Whether the real world applications arrive or not, you can speculate for yourself. You really don't need to borrow the equally unsubstantiated opinion of someone else.
The formal class is called BQP, in analogy with the classical complexity clas BPP. BQP contains BPP but there is no proof that it is stictly bigger (such a proof would imply P != NP). There are problems in BQP we expect are not in BPP but its not clear if there are any useful problems in BQP and not in BPP, other than essentially Shor's algorithm.
On the other hand it's actually not completely necessary to have a superpolynomial quantum advantage in order to have some quantum advantage. A quantum computer running in quadratic time is still (probably) more useful than a classical computer running in O(n^100) time, even though they're both technically polynomial. An example of this is classical algorithms for simulating quantum circuits with bounded error whose runtime is like n^(1/eps) where eps is the error. If you pick eps=0.01 you've got a technically polynomial runtime classical algorithm but it's runtime is gonna be n^100, which is likely very large.
Not to defend Google, but they end up saying much the same:
> The next challenge for the field is to demonstrate a first "useful, beyond-classical" computation on today's quantum chips that is relevant to a real-world application. We’re optimistic that the Willow generation of chips can help us achieve this goal. So far, there have been two separate types of experiments. On the one hand, we’ve run the RCS benchmark, which measures performance against classical computers but has no known real-world applications. On the other hand, we’ve done scientifically interesting simulations of quantum systems, which have led to new scientific discoveries but are still within the reach of classical computers. Our goal is to do both at the same time — to step into the realm of algorithms that are beyond the reach of classical computers and that are useful for real-world, commercially relevant problems.
It would be interesting to see what "standard benchmark computation" was used and what its implementation would like in a traditional computer language.
Thanks. We did check out Tailscale, but they didn't quite have what we were looking for: some high-availability custom component that plugs into a low-level container runtime. (Which makes sense, it's pretty different from their intended use case.)
Modal is actually a happy customer of Tailscale (but for other purposes). :D
So if a company only needs an outbound VPN for their road warriors and not an inbound VPN to access internal servers, vprox could be a simpler alternative to Tailscale?
Exactly. Apart from waking up the vehicle and closing the high voltage connectors, a whole myriad of things run off the low-voltage system, including lights, climate, infotainment and most (if not all) actuators.
reply