643 karma · joined February 13, 2014
> Anecdotally, my hardest problems w/ nl2sql are finding the right tables and adding the right filters.
Totally! especially in large orgs with thousands of tables. Using your existing dashboards and queries, gives useful context on picking the right tables for the query.
When Metabase (or any other tool) builds an assistant, we aim to use it to further extend MinusX's capabilities!
> The new lawsuit filed in federal court in Northern California on Monday says that Altman and Brockman “assiduously manipulated Musk into co-founding their spurious non-profit venture” by promising that OpenAI would be safer and more transparent than profit-driven alternatives. The suit claims that assurances about OpenAI’s nonprofit structure were “the hook for Altman’s long con.”
No way is this in anyway close to 8 doublings though. That would take 12 years or by 2035. (1.59^12 = 261x)
However, if I start with: "What is the earliest record of cheating in the olympic games?" Then all models get the question right. It's surprising that GPT-4 gets it right on the first go.
> According to the U.S. Geological Survey, published scientific estimates of the global CO2 emissions for all on land and submarine volcanos “lie in a range from 0.13 gigaton to 0.44 gigaton per year.”
> This is a fraction of the CO2 produced by human activity. In 2021, the global CO2 emissions from energy combustion and industrial processes alone reached a record high of 36.3 billion tonnes (or gigatons, GT), data from the International Energy Agency (IEA) showed
Follow up (noob) question: Are you using a KV cache? That would significantly increase your memory requirements. Or are you forwarding the whole prompt for each auto-regressive pass?
1. How many GroqCards are you using to run the Demo?
2. Is there a newer version you're using which has more SRAM (since the one I see online only has 230MB)? Since this seems to be the number that will drive down your cost (to take advantage of batch processing, CMIIW!)
3. Can TTS pipelines be integrated with your stack? If so, we can truly have very low latency calls!
*Assuming you're using this: https://www.bittware.com/products/groq/
I'm achieving consistent 450+ tokens/sec for Mixtral 8x7b 32k and ~200 tps for Llama 2 70B-4k.
As an aside, seeing that this is built with flutter Web, perhaps a mobile app is coming soon?
1. Similar to other initiatives (mainly opencompute but also PyTorch, React etc), community improvements help them improve their own infra and helps attract talent.
2. Helping people create better content ultimately improves quality of content on their platforms (Both FoA & RL)
Sources:
[1]Interview with verge: https://www.theverge.com/23889057/mark-zuckerberg-meta-ai-el... . Search for "regulatory capture right now with AI"
> Zuck: ... And we believe that it’s generally positive to open-source a lot of our infrastructure for a few reasons. One is that we don’t have a cloud business, right? So it’s not like we’re selling access to the infrastructure, so giving it away is fine. And then, when we do give it away, we generally benefit from innovation from the ecosystem, and when other people adopt the stuff, it increases volume and drives down prices.
> Interviewer: Like PyTorch, for example?
> Zuck: When I was talking about driving down prices, I was thinking about stuff like Open Compute, where we open-sourced our server designs, and now the factories that are making those kinds of servers can generate way more of them because other companies like Amazon and others are ordering the same designs, that drives down the price for everyone, which is good.
This isn't correct. Here's a thought experiment to clarify:
If we burn X MJ of oil, we transfer X MJ of energy as heat to our house.
Otoh, if we set up a generator (say 40% efficient) which produces electricity from oil and dumps waste heat into our house, and use that to power a 300% efficient heat pump, we get 0.6X (generator waste heat) + 1.2X (300% * 0.4X) = 1.8X MJ of heat using the same X MJ of oil.
Thus,we get an additional 0.8X MJ of heat (80% extra heating) from the same X MJ of oil just by using a heat pump (& generator) compared to burning it. I.e we're using the energy stored inside the oil more efficiently.
Heat always flows from a high temperature to a low temperature. However, we want heat to go the opposite way (i.e. from the cold outside to our warm houses). There's a technique to do this. We use a gas (the refrigerant) to transfer heat.
We first expand the gas (which cools it) until it cools below the cold outside. We then bring it near the cold outside where it now starts absorbing heat until it matches the cold outside temperature. We then move the gas and compress it until it's temperature matches/exceeds our desired warm inside temperature. Then, we bring it to the warm indoors where heat will now flow out of it.
During the entire cycle, the gas is inside a closed loop. It exchanges heat through radiators. The compression & expansion cycles uses energy (technically they can be offset against each other a bit) which is added to the gas (conservation of energy) raising its temperature.
Thus by supplying a little electricity, we're able to move heat from the cold outdoors to the warm indoors. The high efficiencies are because for using X units of energy, we're able to heat the house by X+Y units where Y is the heat transferred from the outside. Typically Y >> X.
To add to this, the typical efficiency is 300% or more. Since typically most of the heat comes from being "transferred" from the outside.
I'm saying this technology is a net positive for me. Sure, problems can arise from applying this net-positive technology but even considering them, they're a net positive. Especially in this case where we already have solutions for large amounts of crap content created by humans today. I think they'll scale well to crap content created by AI tomorrow (eg: Recommendation systems, p2p sharing etc).
Regarding your second point. I think we just have different perspectives on this. I see this technology as allowing me to do "strictly more" than I could before. You say that it's not "my expression" but I say it is because I'm able to bring forth what's in my mind (according to me!). You say it's not what I had in mind but compare against this scenario: I have an idea for a song and use a musical instrument. The output that I produce is quite far from my intent as I'm untrained. With this model, I can get a lot closer to my intent compared to using an instrument since I'm untrained. I still have the option of later learning an instrument if I'm unsatisfied.
Overall, I lose nothing and gain a nice power and still retain the option to learn an instrument if I wish. Thus it's a clear net positive for me that allows me to express more right now.
[0] https://ai.meta.com/blog/generative-ai-text-images-cm3leon/
> As such, and as described in our paper, we’ve trained CM3leon using a licensed dataset.
Regarding whether it is "my" expression or not, I consider expression as a way of making my feelings or thoughts known. This grants me an additional way of making my feelings or thoughts known. Perhaps of lower fidelity than if I learn to play an instrument but maybe that will improve with time. Either way, since I now have an additional choice of expressing my thoughts and _I_ decide if it's "good enough" to publish, I consider it as allowing me to express more.
Re:stable homes, it'd be interesting to envision whether non-nuclear family structures would be more stable. Maybe matrilineal/ matrilocal cultures (even if they're patriarchal) where children of a family are those born to it's women (more stable family since you don't divorce your siblings/mom, much freer relationships, paternity maybe unknown etc). The book Sex a At Dawn (controversial book!) explores these ideas.
1. Su Song
2. Plato
3. Gautama
4. Zhang Heng
5. Bhaskara(I/II)
6. Al Biruni
7. Da Vinci
8. Al Khwarizmi
[1] https://techcrunch.com/2014/03/24/sheryl-sandberg-facebook-r...
[2] https://news.ycombinator.com/item?id=34227388 - Steve jobs email asking Google to stop poaching from apple and them agreeing
Regarding your experience, a lot of advertising is retargeting or keeping the brand alive in the minds of buyers or remind them constantly (since humans are forgetful). This statistically increases the probability of purchases made amongst a cohort (very measurable). Personally, I think this is user-hostile and is a strong case for having more control over the kinds of ads we're exposed to or limit/penalise harmful ones. Taken to the extreme (going on a slight tangent here), if we really owned/controlled our devices and software, which means having control over consumption, we'd block all ads which would incentivise platforms to remove the distinction between content and ads and we'd do more content filtering on the client (This is however computationally inefficient, a sort of arms race).
Regarding influencing behaviour, it's a lot more efficient to match the right seller to the right buyer rather than influence buyer behaviour (which is hard). However, to achieve that extra marginal gains/returns on ad spend, advertisers often do attempt to change buyer behaviour (through building brand associations, retargeting etc). Again, better user control would help here.
Agree on street ads and other public real world ads as there is no consent here (Real world movement is a need, not a choice).
Regarding ads vs word-of-mouth, Ads really are very effective for many people, in a measurable way, both quantitative and qualitative. It of course will differ from person to person.
Regarding increased inefficiency and effect on the economy more generally, ads actually increase efficiency in the system as long as they are not rent-seeking (i.e. pay to jump the queue) or user-hostile. Eg: It helps many niche businesses reach customers which are otherwise very hard to reach (due to geographical or novelty constraints). Taken to the extreme, all value gained from advertising will ultimately flow to the ad platforms. Even in this extreme case, overall consumption increases (due to buyers buying more) as well as competition (due to many more competitive small businesses, innovators dilemma theory) which is good for a capitalist economy (i.e. makes it more dynamic in both choice and inequality) and increases GDP (a poor metric).
As an aside, it's also a progressive tax (I.e rich and poor both consume the same good but advertisers pay more to reach the rich, thus funding the good for the poor).
Some extreme opinions below on the fundamentals:
<Rant>
Even in localist/anarchist/decentralised utopias, ads won't die (especially personalised). It's a fundamental need of humans and the only other form of payment (apart from paying for compute itself) that doesn't depend on a system of violence to enforce it. As long as there exists economies of scale, there will be marketing (to increase consumption and hence cheapening it per capita i.e. use less labor), and hence advertising (to solve the unknown-unknown problem or presenting/impressing).
i.e. Even when we eliminate competitive enterprise, build localised production of needs, and have full control of our devices, if we want to build something cheaply or spread an idea, we'd have to advertise. It would look a lot more like "advertising on the merits" with full user control though considering human nature, there will also be a lot of "advertising focused on presentation".
</Rant>