adamsmith

We have not announced anything formally RE fundraising, but we are well funded and growing the team quickly. All offers include competitive equity.

> it's always more worthwhile to debate and work with people who have to consistently match their predictions against reality

Yes! Fast, clear feedback loops provide such a boost!

If it’s helpful here is the prompt I use to clean up voice-memo transcripts: https://gist.github.com/adamsmith/2a22b08d3d4a11fb9fe06531ae...

For Python's syntax: https://docs.python.org/3/library/ast.html

If you want examples of code → ast, googling for [python ast visualizer] turns up a few tools

There’s a lot to unpack in the quoted passage, but it is not an ad hominem attack. That you think it is, and then turn around to make an ad hominem argument of your own explains your misinterpretation.

If anything this will turn out to be the opposite of the truth: People hit by autonomous cars will have more to go after in court than those hit by individuals, many of whom don't have adequate coverage or savings to compensate victims. The minimum "bodily harm" insurance coverage in California is only $15,000.

Ivan Sutherland wrote one of my all-time favorite pieces, called Technology and Courage. It's a short read, here: https://cseweb.ucsd.edu/~wgg/smli_ps-1.pdf

He talks about the courage it takes to do risky work in our field, and gives practical techniques for overcoming barriers, such as having collaborators, deadlines, "just get started", stock compensation, etc.

Dr Sutherland also discusses the courage to keep going, or even _stop_ working on a project.

Given his broad background, he discusses how these dynamics play out in a wide range of fields, including education, startups, and research.

For me it is always a visceral read.

Thank you Quinn! It's been both cool and instructive to see Sourcegraph take off. Godspeed!

It could be doing some "fine tuning" based on the repo. That would be cool! That said, what I meant when referring to 'understanding' the non-local nature of code was in a more principled way.

For example, if an object defined in another file has a function called `rename` that takes zero arguments, when calling it from another file Copilot will likely suggest arguments if there are variables like `old` and `new` near the cursor, even though `rename` actually doesn't take any, just because functions called `rename` typically take arguments. This behavior is in contrast to a tool like an IDE that can trace through the way non-local code references work.

Does it seem to only understand imports of public libraries? If so, it's likely that, rather than understanding the contents of those libraries, it's learning from others' use of those library APIs. If not, it is likely just understanding the words in the API at a shallow depth.

Yes, this was precisely what I was referring to. In small-enough programs (e.g. one file) Copilot has all the context. The other extreme would be something like the Chromium codebase. Because of this, Copilot looks better in quick demos than real-world use. (Though of course it is very impressive and this tech will get there, hopefully very soon!)

The higher the frequency the more data you can send. This is known as the Shannon Limit. https://en.m.wikipedia.org/wiki/Noisy-channel_coding_theorem

Super interesting question. I don't know.

I suspect they will dominate all the largest technology markets for the long run, unless the government intervenes. I might be wrong.

They may not dominate the future of "technology markets", though, if the sum of the long tail dominates. Right now FAANG is 42% of the NASDAQ Composite 100.

> I've tried to explain this to VC firms. Instead of making one $2 million investment, make five $400k investments. Would that mean sitting on too many boards? Don't sit on their boards. Would that mean too much due diligence? Do less. If you're investing at a tenth the valuation, you only have to be a tenth as sure.

This turned out to be absolutely right, and the < $2M checks have ballooned in volume. They are invested by firms that do not take board seats and do comparatively less due diligence.

That said with the benefit of more hindsight I think the reason there are no more Googles is that FAANG has a stronghold on the largest technology markets.

Yes, there are orders of magnitude of scale in both fields.

In software it can cost $10k to build an app but Google Search costs $10^10: a six orders of magnitude difference.

In music there’s got to be a big dynamic range in effort involved as well. (I doubt it’s 10^6 though!)

But comparing the best music to the lowest end app is apples to oranges.

The turnarounds always happen quickly after the fed takes serious action. Consider the 2008 crisis, or take 1933 (I believe?) when we broke from the gold standard: immediate exit from the Great Depression.

The fed and congress have been very on point with the stimulus response speed this time around and so I think we’ve seen the bottom of the market through this. And most investors are voting with their dollars that this is true.

(I’m just parroting Ray Dalio.)

Hi there, Kite at most about half of what TabNine uses, often less than that.

Thanks for the thoughtful message!

We hope you'll give Kite another shot. We're always improving it and have covered a lot of ground in the last few years.

In particular, a common complaint about TabNine on Twitter is that it's heavyweight. Based on our comparisons of today's Kite versus today's TabNine, we believe Kite uses less CPU/battery, and far less memory. (My teammate dhung posted a longer comment about TabNine.)

We hear you. We agree this was a mistake, and over the ~three years since then we've listened to the community, for example by releasing local processing for Kite.

Thanks, we believe we fixed this. Could you try again?

Thanks for your patience!

Hi there, thanks for the question.

IntelliCode has a similar vision but uses much simpler models. IntelliCode can essentially give you to the top 5 attributes for a global type.

Kite uses statistical models which are orders of magnitude larger and richer, both in terms of the context they understand and the sophistication of the completions provided.

Here is a good example from the blog post:

  var server = express(); server.

IntelliCode returns:

  use, set, get, listen

Kite's first completion is

  use(bodyParser.json

(bodyParser is defined earlier in the file.)

Finally, IntelliCode only returns completions for attributes (foo.) whereas Kite returns deep learning completions in all the places they are useful.

Adam from Kite here. The team and I are thrilled to launch our new deep learning engine for JavaScript completions.

After years of work building our Python product — now to over 250,000 monthly active users — we decided we were ready to expand a new language. JavaScript was a clear choice given how well it pairs with Python, and it was the our most-requested language.

We took what we learned from building our Python product, adjusted it to JavaScript’s nuances, and trained the model on 22 million open-source JavaScript files. We then spent months dogfooding, user testing, and applying advanced model filtering techniques to ensure the UX was ready for everyone to enjoy.

We think the end-result will help you have more fun writing JavaScript, while also helping you ship software faster. We can’t wait to get feedback from you to help us improve our JavaScript completions from here! We'll be online to answer your questions here today as well.

Since a lot of folks ask, Kite continues to work 100% locally and Kite for JavaScript is free. We plan to release support for more languages later this year, and are taking requests at kite.com/letmeknow. Thanks!

Hi, this is Adam from Kite. Thanks for checking us out and for the write-up!

The DGPU point caught my attention. Kite should only use CPU, even for its tensorflow tasks. Could you email me at [email protected]? I'd love to learn more about what you saw.

Thanks again!

You need semantic understanding to do several of those operations. Parsing often isn’t sufficient.

Congrats on the fundraise Greg and team!

Does this mean that OpenAI may not disclose progress, papers with details, and/or open source code as much as in the past? In other words, what proprietary advantage will Microsoft gain when licensing new tech from OpenAI?

I understand that keeping some innovations private may help commercialization, which may help raise more funds for OpenAI, getting us to AGI faster, so my opinion is that could plausibly make sense.

Kite used to be cloud-powered which meant users had to trust us with their code.

In January we announced "going cloudless" -- we moved all processing back to the client and code is no longer sent to our servers. (You can find the blog post announcing this here @ https://kite.com/blog/launching-line-of-code-completions-goi...)

Great to hear Kite has been a useful part of your workflow, thanks for being a user!