williamstein

They are doing a better job than Anthropic (etc); at least it is possible for the general public to use their product…

I'm not answering your question, but I just tried voice input on the ChatGPT iOS app. They integrate their whisper model to directly do speech recognition, and it is always broken for me, with "There was an unexpected error (API.API.APIError, 0)" on any input. Of course, I can also just type the mic and use the normal iOS text-to-speech in the input box, and that works fine.

YES. It's extremely clear and easy, in multiple ways, both through the ChatGPT app and through the Subscriptions Tab in settings in iOS. This is one of those things that Apple does reasonable well, compared to some vendors.

The in-app purchase to upgrade to plus is very smooth…

Thanks! Is there a status page like https://oblivus.com/availability/ for Lambda Cloud? I got very excited about building a product on top of Lambda cloud for cocalc.com, but after read lots of docs, when I tried Lambda cloud, I just got the message: "We are currently out of capacity for all instances. Please check again in a few hours." This was on April 27. I immediately thought: "There is no possible way I can build a product on this for my users, but it may be very useful for other tasks, e.g., training." If there were something like https://oblivus.com/availability/, especially with historical data, it would be very useful to appropriately set my expectations about how Lambda Cloud can best be used.

By the way, Lambda Stack is VERY impressive. Thanks for maintaining that!

I think they mean that almost 2600 are literally available right now for you to rent - or if you requested them, then you would get them right now, and then there would - for a while - be 0 that are available, because you have them all.

I wonder if Google has not yet released an API due to a desire to keep control and also ensure their products (docs, gmail) have a major lead over everybody outside Google? What is their motivation for not providing API access? It can’t be purely due to it being too difficult for them, since they are very skilled engineers.

The landing page says "Easy integration via standard APIs Claude can be incorporated into any product or toolchain you’re building with minimal effort." Then there is a big button "Request Access", which for me right now just does nothing. OpenAI has really faced the pain to make their product available via an API to the general public at scale, but Anthropic/Google/etc. don't quite seem to be there yet. It's frustrating.

Thanks for the clarification about what makes Zep unique. Very cool!

From their FAQ: "How compatible is Mojo with Python really? Mojo already supports many core features of Python including async/await, error handling, variadics, etc, but… it is still very early and missing many features - so today it isn’t very compatible. Mojo doesn’t even support classes yet!".

Overall, pure Python seems to be about 100x slower than what you can reasonably get with a compiled language and some hard work. It's about 10x slower than what you can get from JITs like Pypy and Javascript, when such comparisons makes sense.

I agree that Mojo remind me of Cython, but with more marketing and less compatibility with Python. Cython aspired to be a nearly 99% superset of Python, at least that's exactly what I pushed my postdocs and graduate students to make it be back in 2009 (e.g., there were weeks of Robert Bradshaw and Craig Citro pushing each other to get closures to fully work). Mojo seems to be a similar modern idea for doing the same sort of thing. It could be great for the ecosystem; time will tell. Cython could still also improve a lot too -- there is a major new 3.0 release just around the corner!: https://pypi.org/project/Cython/#history

I can't believe how many distinct "longterm memory for LLM" libraries have launched in the last month or so! It's intense.

https://www.phind.com/ is also similarly good, but for dev related content.

Weaviate is part of the Go ecosystem and has more built around it, e.g., you can configure it so that it will automatically call openai to do vector embeddings for you. Qdrant is part of the Rust ecosystem, and is more focused on the core problem of being a highly efficient (both in time and space) scalable vector database, but with extensive highly integrated support for queries that combine a vector with a query about the payload (basically a JSON object). They both seem like really amazing projects to me. Both are very liberally licensed open source and have a complementary cloud hosted option.

Which is better for you might come down to Go versus Rust, and also whether you want something that feels more like a complete solution (Weaviate), or something that is more focused (Qdrant).

I agree -- I listened to a 2022 podcast interview with the cofounder of Pinecone and he seems very dismissive of what open source can bring to vector database. I found that annoying and off putting. The open source vector database ecosystem is really impressive right now!

I didn't post this, but I coincidentally tried testing numerous open source vector databases during the last week, including building them from source, testing them for my use case, etc. For my needs, and thinking longterm, I think Qdrant is the best option. Their docs are clear and it was easy to build from source using Rust (it takes about 30 minutes, but it just works) and the binaries are relatively small and portable. It was also easy to build the latest dev version, which has built in authentication (due to a bounty claimed two weeks ago). There's a new official Javascript client (for use from nodejs), which builds on "openapi" and seems to work well, if you're using nodejs as a client instead of Python.

My overall impression of Qdrant is that it's clean and uncluttered compared to other options, development is very open, and the devs are technically very capable. The docs are clear and to the point, without a bunch of extra nonsense getting in the way. It's also easy to try out for free via their hosted cloud.

They put my site https://cocalc.com, which has chatgpt API integration, into this gpt4free. As a result, I had to modify https://cocalc.com to require sign in before providing the ChatGPT functionality to visitors, and I also explicitly updated our terms of service to clarify how our API can be used. I made a pull request https://github.com/xtekky/gpt4free/pull/461 to Gpt4free to have them remove cocalc. They were respectful, with some discussion back and forth, and they merged the PR. I personally don't think that Gpt4free should be taken down, so long as they respect the explicit requests of projects they proxy. They were certainly respectful with cocalc.

It sounds like unnatural "podcast speech".

Presumably ChatGPT is better at writing database queries or answering questions about a large Csv file than somebody who doesn’t know how to program. This was the subject of a PR talk last week be an openai cofounder - https://youtu.be/C_78DM8fG6E - my guess is something like that is what inspired this article.

A google image search for "Technical Author" has a lot more women in the results, and also this very article lower down the page...

My company pays its monthly Google cloud bill using a Chase credit card, and we get maximal credit card points because Chase classifies Google as an advertising company. Go figure.

Ilya gives numerous talks and interviews, and he's well worth listening to about technical matters. I listened to many of his talks recently, and the main theme is that scaling up compute works, and will continue to do so. His optimism about the potential of scaling to support deep learning has clearly guided his entire career, starting with his early success on AlexNet.

All but one of the eight authors of that paper left Google, as laid out in this thread: https://twitter.com/Contrary_Res/status/1621203016903954438?... For example, "(1) Aidan Gomez (@aidangomezzz) left google and founded @CohereAI in September 2019".

1. I'm integrating ChatGPT extensively into https://CoCalc.com. This integration makes a lot of sense, because cocalc is a platform in which relatively inexperienced students use Jupyter notebooks, linux terminals and Latex. So far, the most popular feature by far is a "Help me fix this" button that appears above stacktraces in Jupyter notebooks.

2. One software engineering challenges is that ChatGPT often outputs code in markdown blocks. I've had to emphasize in prompts that it should explicitly mark the language. I then got inspired to make it possible to evaluate in place the code that appears in these blocks using a Jupyter kernel, and spent a week making that work (so, e.g., if you type a question into the chatgpt box on the landing page at https://cocalc.com, and code appears in the output, often you can just evaluate it right there). There seem to be endless surprises and challenges though. For example, a few minutes ago I realized that sometimes the giant tracebacks one gets when using Python in Jupyter notebooks are so big (even doing simple things with matplotlib) that they end up resulting in too much truncation: https://github.com/sagemathinc/cocalc/issues/6634

3. I'm mostly using GPT-3.5-turbo rather than GPT4, even though I have a GPT4 api key. Aside from costs, GPT4 takes about 4x as long, which often just feels too long for my use case. The average time for a complete response from GPT-3.5 for my application is about 8 seconds, versus over 30s for GPT4.

This is fantastic. The discussion and content so far are interesting and high quality. Thanks.

The debug library https://www.npmjs.com/package/debug is incredibly popular for Javascript logging, and is also incredibly simple (just a few lines of code).

I'm interested in Beam Analytics partly because I personally find the new Google Analytics 4 user interface (especially around cohorts) to be surprisingly frustrating to use, and I have used the previous versions of Google Analytics for 10+ years. Is it just me, or did Google Analytics 4 also just get redesigned in a strangely more frustrating way?

GPT's output is by default somewhat random. If you ask the same exact question several times, you'll potentially get several different answers. Each successive word in the output is chosen from a distribution of possibilities -- that distribution is fixed, but that actual sample chosen from the distribution is not fixed. See, e.g., https://platform.openai.com/docs/api-reference/completions/c...

If you're doing software development, then https://www.phind.com/ with the expert checkbox uses GPT-4 and is better than the GPT-4 subscription, and is free. I just learned about this on Hacker News yesterday: https://news.ycombinator.com/item?id=34884338

I just did a little test involving a question about the Link component in Nextjs, and phind is aware of the major API change nextjs recently made to that component, whereas of course chatgpt/gpt-4 isn't, due to the training data cutoff.

I would not discount the potential of LLM's to support research level mathematics. There's interesting research going into combining the LEAN theorem prover with LLMs, e.g., this recent paper: https://arxiv.org/abs/2202.01344

"... automatically solving multiple challenging problems drawn from high school olympiads."

This combination of GPT and LEAN is the first thing that has finally got me really interested in the potential of LEAN. And LEAN very directly addresses the "bullshit" issue, since it formally verifies whether or not a proof is correct (it's like a compiler like Rust, but for mathematical proofs).

I am glad it tokenizes Python and all other programming languages in a systematic way.