dweinus

If I understand the paper right...

At indexing time:

- run LLM over every data point multiple times ("gleanings") for entity extraction and constructing a graph index

- run an LLM over the graph multiple times to create clusters ("communities")

At query time:

- Run the LLM across all clusters, creating an answer from each and score them

- Run the LLM across all but the lowest scoring answers to produce a "global answer"

...aren't the compute requirements here untenable for any decent sized dataset?

...don't forget about Planetary Annihilation! Not free or open source, but still a lot of fun! https://planetaryannihilation.com/

If 80% of employees stay each year, then the expected tenure would be about 3 years.

Genuinely curious: do any of the features/use cases they market here sound appealing to people? I am struggling to imagine wanting to use any of them. I can, however, easily imagine the sound my cooling fan will make while the NPU is running 24/7.

100% accurate. Impressive

This.

"Since 2014, it has raised its menu prices by 100%"

... during the same time period, US inflation has been 32%.

Carbon capture from the air is at best a tiny fraction of the solutions we need. Right now our time is better spent on emissions reduction.

https://www.nature.com/articles/d41586-023-00953-x

Cool, but carbon capture from the air is at best a tiny fraction of the solutions we need. Right now our time is better spent on emissions reduction.

https://www.nature.com/articles/d41586-023-00953-x

Thinking out loud... I don't think these problems can be solved. If you are going to do it anyway, I would suggest:

- Using a RAG architecture on top of a database of factual information. Wikipedia is probably your best bet. It is not 100% factual or correct either, but maybe as good as it gets. Scaling RAG to wikipedia size is not trivial, but I think it can be done.

- Prompting the LLM to cite its sources so people can fact-check the fact-checker

- Prompting the LLM to say it is unsure when something does not have a clear answer. I don't expect this to be reliable, but maybe somewhat better

Andrew, you and your friends should be proud. It's really encouraging to see people, especially in your generation, thinking seriously about the problem of misinformation. There are fundamental challenges you all will face in this idea:

- most current LLMs are trained on large amounts of web data that itself contains facts, opinions, and misinformation. These things are treated equally, so I would expect the LLM to get common facts right, but also to represent opinions or misinformation as facts when they are pervasive.

- LLMs "hallucinate" and tend not to know when to say "I don't know" or to not try to fact-check something that is not factual in nature.

...in short, I would expect LLMs to be an unreliable fact checker, which has the potential to do as much harm as good.

This paper is bad and Gizmodo should feel bad.

Besides making ridiculous and unsupported causal claims (as other commenters have noted), the results are really poor (R score of 0.2). They show that the model achieves human-level performance, which is also really poor. Then they run it on natural images of politicians and the results are significantly worse. What the paper actually shows is that facial appearance can not consistently predict political alignment, but the authors want to try really hard to do it anyway.

Yay, the planet is saved! /sarcasm

What would these tabular foundation models look like? LLMs work as foundation models because the input is fixed in format (a sequence of text). Would the model be for a specific fixed tabular format?

> Your city has a problem with trash

...doesn't that entire problem rest on the fact that the unstated goal is to pick up the trash at the lowest cost labor will bear? That yields service reduction, automation, and labor disputes. When you look at it that way, the organization is seeking the goal, not fighting it. I think there is truth in the general idea, but a correlary that no one sees value in continuing to solve a problem that has been solved once. They always imagine it will get cheaper, meanwhile those who create expertise in it see themselves as more valuable with time, and that creates tension.

Right. Maybe it is hard to prove, but the choice to eat them is a life-or-death decision. I am going to opt for tofu. Given that "sentience" is roughly the ability to have feelings, it seems very hard to believe most animals are not sentient. Spend a day with a cat or on a farm and you will witness emotions. Do octopuses have feelings? The paper is pretty confident: "There is very strong evidence of sentience in octopods". It seems likely to me that animals that are distant from humans are harder for us to understand, but still sentient.

I did, that's where my quote is from. The appendix confirms; they ran the regressions on just two inputs: 2022 and 2023 totals.

The methodology seems to be: compare commit activity from 2023 to prior years, without any idea of how many involve Copilot. Then interpret those changes with assumptions. That seems a bit shakey.

Also: "The projections for 2024 utilize OpenAI's gpt-4-1106-preview Assistant to run a quadratic regression on existing data." ...am I to understand they asked gpt to do a regression on the data (4 numbers) rather than running a simple regression tool (sklearn, r, even excel can do this)? Even if done correctly, it is not very compelling when based off of 4 data points and accounting for my first concern.

Only a mathematician could think that when people are arguing over a cake, the best solution involves giving everyone a knife (Stromquist moving-knives procedure)

I agree, Facebook is generally lacking in intelligence.

Roughly -8 years from now: https://www.smithsonianmag.com/innovation/inside-first-solar...

We need solutions, likely a mix of them, because they all come with challenges:

1) Transmission takes time and $$$ to build (especially to get the land). Maintenance ain't cheap either. Longer transmission == greater losses.

2) CSP requires a specialized site and equipment, with limited generation/storage and greater losses

3) EVs as batteries to back up the grid externalizes battery wear to car owners. To compensate owners would probably cost more than just building dedicated battery back up on the grid (which would be much more reliable). This idea has never made sense to me.

4) Wind is a great resource, but also variable. There will be times when the wind doesn't blow at night.

...and this is why we can't have nice things

Of course they don't! But I think the most fascinating and exciting part about LLMs is: a sufficiently large model can produce things that look a lot like cognition, without having it at all. That is shocking and suggests maybe AGI is not even goal worth hitting.

Obscure words are excellent when they add more nuaced meaning, but in most situations the same few words are perfectly cromulent.

I was disappointed I couldn't actually create recursion by tracing my egg to a chicken to an egg to a...

That's a good explanation, but problems with silently reducing quantity are: - It attempts to trick the customer - As a customer, it makes it harder to depend on your product or buy predictably (a box of cereal used to last me through the week, now suddenly it doesn't) - It is now harder for me to comparison shop because I need to calculate cost per volume/weight - It is often done at a rate higher than inflation

This site is great, people need more transparency and companies need to be called out

> There's evidence that the earth is warming, but only if you squint a certain way.

Get real. I used to sympathize with global warming deniers. It's 2023. The debate has been had for 40 years and now things are literally catching fire. There is not room for debate anymore.

Exciting! What is the argument for Polars on clusters instead of Dask?

Sounds good, but enforcing it how? The only party here that can do the enforcing is the Google/Apple/Microsoft App Store. I'm not sure they all have the right incentives to expend the effort to enforce that in a consistent and user-friendly way. Otherwise would have already.

> consider instead of using fancy and scary mathematics like “neural networks”, it is just tabulating and counting keywords. Is it still infringement?

I've seen this analogy before. A NN, especially an LLM, is very different in method and outcome. If counting keywords allowed me to replicate copyrighted material, it would probably not be fair use. IANAL, but I imagine the fact that a NN can replicate or compete with the original work makes a difference.