sberens

What happens if max_tokens cuts the model off from generating valid JSON?

Can someone ELI5 the argument for the bill?

My impression is it's something like "news websites provide content that creates engagement which drives ad revenue, and the news websites want a piece of that revenue."

Is my understanding correct? Also, I can see how it applies for Facebook, how does it apply to Google?

link to the study: https://www.sciencedirect.com/science/article/pii/S004313542...

The whole abstract sounds like marketing copy from those websites that don't tell you what their product is:

> By avoiding high depth, shortcuts, and intricate operations like self-attention, VanillaNet is refreshingly concise yet remarkably powerful.

> The sum total of research into this "field" is a series of long chains of philosophical leaps that rapidly escape any connection to reality

Well, 10 years we had just gotten image classification barely working. Now, we have something with reasoning abilities close to the average human. Is it really a philosophical leap to imagine in another 10 years we will have made significant further progress?

5 minutes ago I copy pasted a 200 line react component and asked ChatGPT to explain why I was getting a certain bug. Although it didn't quite get it right, it allowed me to quickly identify the cause.

It only needs to save 10 minutes of my time a month for it to be worth the $20 - I would estimate it saves hours.

As someone more familiar with js/react, is there an equivalently mature reactive frontend framework?

From the article:

> Apple also requires its employees to sign nondisclosure agreements, which forbid the signees from discussing secret projects with unauthorized people. That includes other Apple employees — and most certainly anyone outside the company.

As a reminder, the people worried about AGI are not worried about GPT-4.

They see the writing on the wall for what AI will be capable of in 5-10 years, and are worried about the dangers that will arise from those capabilities, not the current capabilities.

Interesting, about 95% of your comments in the last year are saying something negative about Meta/Zuck.

Hm, my tech job tracker[0] shows that it's probably one of the worst hiring markets ever. Where do you see that the market is still strong?

[0] https://www.hnhiringtrends.com/

I think the fundamental problem with OS-level ads is that they're by design irrelevant. I don't mind Google search ads because I'm searching for something related.

But in the OS, if I'm opening the start menu, my intention could be anything. Microsoft is then forced to serve the most generic ad possible, which is going to be irrelevant 99% of the time.

> Autonomous cars in San Francisco made 92 unplanned stops between May and December 2022

So about one unplanned stop every 3 days? Seems a bit of a stretch to call it "clogging" San Francisco.

> other models are in the ballpark

How true is this? From playing around with Bard and Claude, GPT-4 seems to be significantly better, especially around code generation / understanding.

Author here, happy to answer any questions about the experiment!

> It would have increased profitability at that time, too.

Not necessarily. The profit(headcount) function generally varies over time.

Just to note, 3 years later they did release a phone: https://en.wikipedia.org/wiki/HTC_First

Personally, I'm really excited for the era of generative AI powered customer support.

Hopefully it will bring low wait times, knowledgeable "representatives", and fast escalation.

There was hype around Crypto because people were making large sums of money from it, despite unclear "real" uses beyond easily moving large sums of money around.

There is hype around AI because hundreds of millions of people (including me) are using AI tools every day to help them with their work. In contrast to crypto, random people are not making enormous amounts of money; there are no 18 year olds in AI stumbling into hundreds of millions of dollars. Even OpenAI is not particularly profitable and is mainly sustained through its close partnership with Microsoft.

I think this reversal of the money <-> utility relationship makes these technologies fundamentally different, and lumping them together into the category of "just another hype cycle" is unfair.

> The legislation will leave fewer than 10 big banks in the United States subject to stricter federal oversight, freeing thousands of banks with less than $250 billion in assets from a post-crisis crackdown that they have long complained is too onerous.

> Once the bill is signed by Mr. Trump, small and medium-sized banks will no longer be required to undergo “stress tests” aimed at measuring their ability to withstand a severe economic downturn.

Would the current situation have been prevented/mitigated if not for the repeal?

Oops I meant text embeddings

I'm looking forward to an OpenAI quality embedding that I can use locally. I think eliminating the latency/cost from calling their API will allow for some novel UX patterns.

> As WebGPU does not support FP16 at this moment, the memory consumption of running the demo is about 7GB. For Apple silicon Mac with only 8GB of unified memory, it may take longer (a few minutes) to generate an image. This demo may also work for Mac with AMD GPU.

Ouch. It will be interesting what this number will look like in 2-3 years though.

Yeah it only counts top level comments

Ah thanks for pointing that out! I'll see if I can fix it.

Edit: should be fixed!

Actually I just realized because the "looking" comment format includes "remote: ", filtering by "remote" doesn't do much. You should filter by "remote: yes" if you want to see people who want to be hired remote.

I was thinking of preventing this using semantic search with OpenAI's embeddings, but I realized that the total size of the embedded comments would be around 2GB, which most people probably do not want to download.

Thanks! Yeah, I'm working on adding an option to normalize everything by some measure of activity around each data point.

Really cool! FYI, the labels in the x-axis for the first two charts overlap with each other for me. https://pasteboard.co/4ts0uaUQopxJ.png

I guess it's time for me to put prompt injection attacks into my submissions

I think yes, at least for coding. Pre-ChatGPT, when I was working with an unfamiliar language or framework, my workflow would be: google something -> skim top links for what looks helpful -> click on one of them -> parse through mountains of SEO'd text -> get the answer.

The workflow I have now with ChatGPT and what I image it will look like in the new era of search is: query -> read result written for a human, not a search engine -> (10% of the time) check if the result is hallucinated.

Especially for basic questions where I know what the answer should look like, I'm really enjoying the new workflow.