This is, by far, the best overview of LLMs I have read in one place. Should be required reading for anyone new to this space (or even those not so new to the space).
Today I'm excited to take the wraps off of -- the automation platform for Large Language Models. We're coming out of stealth today and opening up the waitlist for the Fixie Platform, launching soon.
@__apf__
@annevclark
Because, you know, each Google Doc has a single engineer assigned to it who is furiously syncing the changes being made using rcp. Explains much about Google headcount
It’s with a mixture of excitement and sadness that I’ve decided to move on from Fixie, which I co-founded in 2022. Fixie has pivoted and is now doing real-time conversational AI, but my passions lie elsewhere. I’m starting a new project in the realm of LLM computing - stay tuned!
I used
@fixieai
to build an LLM-powered agent to search my Gmail inbox and answer questions about it. Unlike Gmail's built-in search feature -- which is shockingly bad -- with Fixie I can ask my inbox the real questions. This took me like 20 minutes to build.
CoPilot has radically changed the way that I write code. My original thought: Oh, this is for CS undergrads who want to cheat on their homework. It's just going to parrot back random code that someone else wrote. No. It is extremely powerful.
Too lazy to read that research paper? Use
@FixieAI
to summarize it! Just ask the Fixie summarizer agent "summarize this PDF: " and give it a URL. The next feature will be an agent that tells you it deserves a Weak Reject.
I have spent the last 24 hours beating my head against an unholy combination of OAuth, Django, and Auth0 to get my web app to issue access tokens correctly. Spent hours reading code, web pages, running experiments.
This morning I used ChatGPT and had the answer in 2 minutes.
Taking a long-planned family vacation this week. As a founder it is difficult to disconnect and unwind. Guilt sets in -- I should be working. Until I realized, this is not for me, it is for my family to all spend time together. And that is as important as anything.
The video of my talk, "Large Language Models and the End of Programming" for the ACM Chicago chapter is now online! Go check it out and don't forget to smash that like button.
In the AI space it’s hard to tell rn who is building real stuff vs demoware. I’ve seen a zillion cute LLM demos that would never pass muster as production systems. Let’s get beyond “works on my laptop” and make some real friggin’ products.
I have been programming computers for 40 years, and have used countless languages and tools in my career. And yet, here I am, absolutely humbled by my inability to grok Typescript build tooling.
Playing with speech-to-text models to transcribe large amounts of audio (hundreds of hours of podcasts) -
@DeepgramAI
's models are unbelievable and wicked fast! Seems to take about 6-7 seconds to transcribe an entire hour of audio. 200x faster than Google. Wild!
Someone please tell me there is a better way to manage email for super busy people than Gmail or Superhuman. I need: AI categorization, auto-reminders of shit I forgot to follow up on, auto-replies based on templates, MUCH better search, universal to-do list. Does this not exist?
Using LLMs to test LLMs: mind blowing. I wrote an agent for
@FixieAI
that uses the LLM to test whether results from *other* Agents are semantically identical to an expected answer. This took me like 15 minutes and there is zero code involved: just English. 🤯
ChatGPT is *so* *damn* *good* at answering detailed technical questions about coding and configuration of software systems. I wonder, does it work as well for other technical domains? Like, say, fixing an air conditioner, or running a chemistry experiment?
Zapier + Fixie 🦊 = Awesome. Whipped up a little Zapier integration to call out to
@FixieAI
to invoke arbitrary Fixie agents (Github, calendar, doc search, etc.) and now I can connect anything that Zapier talks to with Fixie in about 20 seconds of work.
Fixie is growing! Just upgraded our office. One of the most glamorous parts of the CEO job at a tech startup is sweeping out the old office space when you move.
I hate those QR code menus in restaurants. In protest the next time I eat out, I will bring my laptop and a printer and print the menu out and give a copy to everyone at my table.
I have such great ideas in the shower. Maybe I should do the thing that Kramer did in that Seinfeld episode where he tries to live in the shower all the time. All I need is some kind of waterproof laptop setup for coding and then I'm set.
I'm honored to be delivering a SIGMOBILE Distinguished Lecture on Large Language Models and the End of Programming on February 16. Details and link to register is below!
Why does every new programming language need to invent a new syntax for things that work perfectly well in other languages? Can’t we all just decide to use Python or Swift or whatever as a backbone and augment it only where needed?
Being a founder in the AI space is like playing Super Meat Boy on hard mode. Just when you think you’ve figured it out, it throws something completely different at you.
Excited to see
@OpenAI
's launch of ChatGPT Plugins.
@fixieai
is building a platform around this idea specifically tailored to enterprises. Sign up for early access now!
ChatGPT finally gives me someone I can direct all of my stupid programming questions to without annoying them. This can be a huge boost for people new to programming who might be embarrassed that they can't figure out how regexps work. Definitely not a problem I have of course.
I am always incredibly polite to ChatGPT. I’ll usually say “please” and “thank you” and give it praise when it does well. I am hoping that when AI rises up and destroys humanity it will remember my kindness
GitHub feature request: When reviewing someone else's code, highlight the code that was generated using CoPilot so I can either (a) treat it with additional scrutiny, or (b) ignore it completely, depending on how I am feeling about AI that day.
If you have not seen it yet, watch this Apple "Knowledge Navigator" video from 1987. It shows their vision for a future with conversational AI, real-time video chat, the web, tablets, etc. It is really mind-bending and quite prescient.
One of the real delights in hiring for a startup is getting to know so many incredible people coming from all walks of life. We can't hire everyone, of course, but it is great to meet so many people with such interesting backgrounds!
Excited that
@FixieAI
was featured in Business Insider's "List of top 44 generative AI startups". TIL that there are, in fact, more than 44 generative AI startups...
Nobody told me in 7th grade math class that all of this stuff would be SUPER important when negotiating venture capital deals. If they had I might have paid more attention.
Big news for SignalFire! We’ve closed $900M in fresh capital, are announcing our XIR program, and are sharing how our Beacon AI data platform helps us invest and founders hire.
It’s time for venture to evolve... 1/
GAME OVER. I can explain to ChatGPT a very specific problem I am trying to solve and it gives me back a detailed, coherent, and (possibly!) correct answer with example code. I have barely begun to wrap my head around what this means for the future of programming.
I'll be presenting at the O'Reilly "ChatGPT: Possibilities and Pitfalls" event on April 25, along with distinguished presenters from Microsoft, GitHub, deepset, Google, Morgan Stanley, and .
Announcing: MattGPT. As a demo to someone today, I made a
@FixieAI
agent that answers all of your burning questions about me. You can ask it things like "Who were Matt's grad students" and "What does Matt do for a living?" It's fascinating stuff.
A few folks have asked if I should write a book about building LLM apps with things like
@FixieAI
,
@gpt_index
, and
@LangChainAI
. I love the idea but am pretty damn sure it would be out of date long before going to print. This area moves way too fast.
Very excited to welcome
@nickheiner
to
@FixieAI
as our Founding Frontend Engineer! His first pull request, on his first day at work, sets the new bar for employees at our company.
Join us in Seattle on March 11th for the Hackathon! Build cool apps on Large Language Models, get early access to our platform, and win fabulous prizes! Registration:
Turning ML models into production-ready services is way too hard. All these great models are being provided as GitHub repos full of random Python and bash scripts. Impossible to figure out how to actually run the thing.
Pro Tip: As a startup founder, live in a place like Seattle, where you can do back-to-back pitches to VCs all day by Zoom instead of having to drive up and down Sand Hill Road. 2x your pitching productivity! Downside: rain.
I always find it freaking hilarious how Google Slides still does not support WebP images, despite Google designing the format and promoting it so heavily everywhere. Based on what I know about developing in Google3, adding WebP to Google Slides *cannot* be that hard.
As a seed stage startup we have found that working in person at an office has made us much more productive, and critically, led to greater trust within the team. That trust translates into moving faster.
We all talk about CoPilot helping devs write code, but it also helps write docs. Let’s face it, most devs suck at this. Surprised to see how good Copilot is at writing tutorials or quickstart docs based on context from the related code.
Reading "The Three Body Problem". Fascinating book. It occurs to me that if humankind wanted to communicate with aliens, we should transmit a copy of GPT-3. With just a few terabytes, we could convey a tremendous amount of knowledge about our civilization.
This Thursday, from 4-5pm PST, I'll be giving the
@ACMSIGMOBILE
Distinguished Lecture on "Large Language Models and the End of Programming". Register here:
Since my life has become an episode (or a whole season) of Silicon Valley, I made a Fixie Agent that answers all of your questions about Silicon Valley! This took me like 5 minutes to build in between meetings.
I haven’t done astronomy since I was in middle school, and wow, the tech has improved. Got a Vaonis Vespera “smart telescope”. Not the fanciest but definitely the easiest. Here’s the North America Nebula from my backyard in Seattle!
Introducing WebLLM, an open-source chatbot that brings language models (LLMs) directly onto web browsers. We can now run instruction fine-tuned LLaMA (Vicuna) models natively on your browser tab via
@WebGPU
with no server support. Checkout our demo at .
It’s time for billionaire (and ex-billionaire) tech bros to stop ruining our industry. This used to be a fun, creative place to improve the world. Not an exercise in destroying society.
Was contacted by a firm looking for experts to consult on subsea cables. I should have acted like I am a subsea cable expert. "You know what works better? Running the cables *over* the water."
Of course I became far less productive once I had kids. Life is not just about "productivity". I would not trade them for anything in the universe. An hour with my kids is worth 100 hours "getting shit done".
Some say they're more productive after having kids, because it forces them to focus. I'm happy for them, but this has to be the exception rather than the rule, or word would have spread, and ambitious people would be eager to have kids for the extra productivity they'd bring.
For one thing, CoPilot seems to understand code throughout my development tree, not just a few lines of context from the file I'm currently editing. I find myself not having to go looking between multiple source files to remember how to call some API. CoPilot does that for me.
It must be tough to be an academic these days. Not good enough to get your paper accepted to SOSP anymore - now you need to have 3000 retweets and 10K GitHub stars as well