Alistair
@AlistairPullen
Followers
3,283
Following
745
Media
80
Statuses
992
@cosineai • YC W23
London, England
Joined January 2014
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
#KAZZAWARDS2024
• 540610 Tweets
SAROCHA REBECCA IN KAZZ
• 456964 Tweets
NORAWIT WITH PAUL SMITH
• 126997 Tweets
Varane
• 37744 Tweets
インプレゾンビ
• 33691 Tweets
魔法少女
• 25290 Tweets
B1NI TOPS SPOTIFY
• 24610 Tweets
FOURTH x KAZZ 2024🥳
• 23554 Tweets
#Destiny
• 18211 Tweets
#うたコン
• 16437 Tweets
名誉毀損
• 14118 Tweets
モグコレ
• 13166 Tweets
#仰天ニュース
• 12828 Tweets
風の行方
• 11763 Tweets
いなば食品
• 10216 Tweets
Pinned Tweet
It's been a while since, but I'm super proud to show off
@cosine_ai
today.
This AI knowledge engine can tell you how to do tickets, explain disjointed and complex features, and find code by functionality.
You can try it in-browser here:
9
10
80
So I've been working on something.
Introducing Buildt: Google for your codebase! Our AI-powered search allows you to find code by searching for what it does, not just what it is.
Available now as a VS Code extension for JS and TS projects, with 15 more languages coming soon!
60
118
1K
@PodsDank
I’m an iOS dev who felt nostalgic after watching your videos on the iPod classic, unfortunately I don’t have mine anymore so I made my own - with fully functional clickwheel!
41
43
574
I remember early this year I was having an issue with our product and I pinged OpenAI support at 8am on a Sunday, and
@gdb
personally helped debug it with me over slack for nearly an hour, truly next level stuff.
stories of gdb’s superhuman abilities from people who worked with him are wild. like when gpt4 first finished training it didn’t actually work very well and the whole team thought it’s over, scaling is dead…until greg went into a cave for weeks and somehow magically made it work
32
136
4K
6
19
608
Demo day today, I’m presenting in around an hour or so, wish me luck
28
14
458
I've made a ChatGPT plugin for Buildt! 🚀
We've had a lot of requests to create an interface that allows non-technical team members to understand and ask questions about the codebase without going through engineering
This plugin allows you to do that:
10
35
380
I’ve had early access to the ChatGPT API for the past couple of months, here are all of the learning’s I’ve found that have made it into our product
30
47
336
So it turns out GPT-4 can write you a perfectly good vector db index in pure TypeScript in about 25 minutes which performs pretty well.
This was a query across 100k vectors of 1536 dimensions:
11
25
233
So after what feels like an eternity of fixes and improvements, here's what we've been cooking up:
🔍 Overhauled search so it gives much more accurate results
🤖 You can now chat with context of your entire codebase
📝 Buildt can now make code changes, and explain code to you
11
18
145
Really great day reaching the final at the
@scale_AI
hackathon, didn’t end up placing top three in the end but was a blast to build and ship something so fast!
6
3
121
What's hilarious about LLM prompting is how helpful memory injection can be – I just improved my prompt's performance on long context examples by simply injecting a line of:
"[Model]: Recalling original instructions..."
Before I ask it to give its answer
7
11
100
Really hit home why the
@ycombinator
process is so great today.
We had to prepare a two-sentence blurb, and pitch it to our partners. We did so and were told ours needed work - too much of a customer pitch.
Within minutes one of the partners helped us create one which we loved!
7
6
86
After a stressful week it was nice to unplug today, and explore a bit more of CA (and get the drone out!)
5
6
80
Thanks
@paulg
for the explanation of what a running back is for the Brits in the audience. Great night at the
@ycombinator
event last night!
2
4
80
Great evening last night at the
@ycombinator
batch meetup, shared some of our issues we’re facing and got great feedback - just the trivial case of putting it into action now. (Yes that is
@paulg
’s head in the bottom left)
4
4
70
This afternoon I realized I needed some kind of basic UI to label code for a dataset we're working on, so I got ChatGPT to build it - took 30 mins. I'm blown away.
3
5
60
To anyone who wants to download the project for themselves here’s a GitHub link to it - lmk if you do anything cool with it 😎
3
6
52
@paulg
As a Brit in SF for the past month the truck thing baffles me, I never see them carrying anything and they take up an inordinate amount of space on the road
13
1
53
For those interested, here are my latest learnings on embeddings, and how they made us overhaul our search engine:
10
5
53
I've been playing around with this idea recently: I needed my completions to be more varied, and even with tempereature=1 I was getting fairly repetitive results. I then intro'd this into my prompts (new UUID on every call) and it's really helped with variety - anyone else tried?
7
3
55
If you're interested, the article was on some tricks we've used for LLM mimicking to reduce our costs and improve latency using synthetic datasets:
6
4
47
This is just the start, we're going to be shipping something soon which takes search even further: have your codebase explained to you. Imagine having a complete expert on your codebase who knows how everything works at your fingertips:
2
3
45
We just went live on Product Hunt, would really appreciate it if you could go and give us an upvote
8
7
44
Met a great founder yesterday who’s working on an alternative meat company (with a cool twist) who said that VCs had genuinely been asking him if his product involved/could involve AI/LLMs - utterly hilarious
4
3
42
I completely rewrote how our search works over the weekend, and it's so much better than it ever was before – by an order of magnitude. It really seems to 'understand' what you're after now, whereas it was a bit iffy on this before, now it has that 'AI' feel.
6
0
41
I’ve been asked to take a look at a few
@ycombinator
applications for the W24 batch — if you’re applying I’m more than happy to take a look, and if you’re not applying you should!
11
2
40
Why would you ever search for code any other way, there isn't a regex out there that could reliably get you these results.
I love being impressed whilst dogfooding
2
2
40
As soon as Twitter went down our signups basically froze, crazy
1
2
34
This was such a nice email to get, puts everything into perspective
1
0
38
With all the hype around LLM agents, I wanted to see if you could actually use one *usefully* in a production environment so I've been working on a specialized agent to figure out what a software project does by giving it filesystem access.
I had to create a fine-tuned Curie…
3
5
36
Had a great conversation with
@WholeMarsBlog
on the way to SFO yesterday – chatted self driving, our upcoming AI release and more!
Yesterday, my Tesla dropped off Cosine AI founders
@AlistairPullen
and
@SamOfStenner
at SFO airport.
We chatted about their product, AI, self-driving, and the experience of going through Y-Combinator.
Watch in 4K:
45
45
316
3
2
34
One of my favorite side features I didn't think I'd use as much as I do in Buildt is the ease of refactoring code like this, it's boosted my productivity a bunch - particularly as you can go back and forth with BuildtBot to get the results you need
1
3
35
Python this week, Java and Ruby shortly after!
Google for your entire codebase in seconds with
@AlistairPullen
's new `Buildt` VS
@Code
extension! 🔥
An
#AI
-powered search that allows you to find code by searching for what it does, not just what it is.
#JS
/
#TS
only, 15 more languages coming soon!
🔗
27
220
1K
5
4
33
Great first couple nights at the YC retreat, already got some great learnings and feedback on what we’re doing. The community is so great
1
0
34
For those interested here’s our alumni demo day clip, today was basically the same as this
4
0
32
Shipping things super quickly that don't scale works great
Until it doensn't
4
1
31
One of the many great things about
@ycombinator
is the constant reiteration about learning from users – everything in this update we've just shipped has been informed by user feedback!
So after what feels like an eternity of fixes and improvements, here's what we've been cooking up:
🔍 Overhauled search so it gives much more accurate results
🤖 You can now chat with context of your entire codebase
📝 Buildt can now make code changes, and explain code to you
11
18
145
1
9
31
Building a good chatbot that can actually execute actions outside of its sandbox and then utilize those results usefully is way harder I thought it would be.
Follow-up questions about prior context can be super hard to manage if you've locked down specific arbitrary actions
3
0
29
@bentossell
When AI can solve webpack issues then I'll believe it's a competent web developer 🤣
1
0
26
The dilemma of trying to come up with a smart solution to a context window problem or just waiting for GPT 4 to drop on Thursday
3
1
28
1
1
26
For those of you interested, here's the repo: - it's very rough currently but I'll be adding some fairly niche stuff over the coming weeks and months (like temporal sharding/sub-indices etc)
4
0
26
Trying to manage both a Hacker News launch and a Product Hunt launch on airplane wifi at 36,000 feet isn’t the walk in the park I thought it’d be!
3
0
25
Well we got to this stage with Buildt much sooner than I expected
5
0
25
I've always wanted Buildt to become more proactive, and I've been playing around over the past few days to do this.
Here you can see an example where it pre-empted what I wanted to search for before I even searched because it could see I was looking for it!
4
2
24
Just published the repo on Github for those interested, gonna try to add some interesting stuff like temporal sharding of indices/sub-indices
So it turns out GPT-4 can write you a perfectly good vector db index in pure TypeScript in about 25 minutes which performs pretty well.
This was a query across 100k vectors of 1536 dimensions:
11
25
233
1
2
24
ChatGPT is much more competent at labelling training data than davinci-003, I’m assuming that it’s thanks to all of the RLHF-ing?
3
0
24
I think this is a great opening usecase, I could see it being very useful for Product Managers or even people in marketing to get a better understanding of how the codebase does things
3
0
23
If anyone else’s in SF for the YC reunion would be great to grab a coffee
2
0
23
So it turns out the most annoying and hard-to-reproduce issue on our backend was down to the GPT tokenizer library we're using to estimate the lengths of context windows – the absolute last place we'd have thought to look 🤯
2
0
22
@alexgraveley
I worked on this for a number of weeks, tried a few things such as deliberately overfitting QLoRA adapters on my data etc but no great results vs in-context, closest I came was using an extended memory transformer - the one I used was this one:
Extended Mind Transformers (EMTs) are a new approach to working with very large contexts and external data sources developed by
@KlettPhoebe
,
@thomasahle
, Normal's AI team. Inspired by the Extended Mind Thesis, we modify Multihead Attention to directly query a vector database.
5
43
221
2
2
21
@dunleavy89
@RenderToken
@akashnet_
Can confirm search assisted by LLMs costs a fortune, you should see our OpenAI bill for the past week 😟
2
1
20
My cofounder just described my job as 'playing with GPT-3' and honestly who wouldn't want that as a job
3
0
21
I can train a model, but when it comes to properly using useCallback in React, I'm totally lost
2
0
21
Also shoutout to
@pie6k
and screen studio for generating the video in the first Tweet, amazing app.
1
1
17
I just used GPT-4 to plan out one of the large core features we have on our roadmap, and it probably saved a good few days of ideating and MVPing
The knowledge cutoff being a decent while ago is a little frustrating at times though
4
1
21
We did it! This search update has been in the works for a couple weeks and I'm delighted to finally ship it.
In short - Buildt's search is now 37% improved in our F1 benchmarks, so snippet retrieval will be even better.
I've been so impressed benchmarking and dogfooding this!
3
1
21
So I just found that the new gpt-3.5-turbo-instruct is ~34% faster than the equivalent chat model on average in terms of TPS!
Ran a test of generating 100 short stories on different topics on both models with the same params and the instruct model was consistently faster.
3
3
21
Code optimisation is very satisfying when you see the execution time consistently drop
0
1
20
spent months writing a language agnostic LSP implementation to make pretty spiderwebs
3
1
22
@mattshumer_
What a happy coincidence
So I've been working on something.
Introducing Buildt: Google for your codebase! Our AI-powered search allows you to find code by searching for what it does, not just what it is.
Available now as a VS Code extension for JS and TS projects, with 15 more languages coming soon!
60
118
1K
1
0
20
It sometimes feels like no matter how much chain-of-thought reasoning you do with GPT-3.5 you still get really dumb answers, then you try it with GPT-4 and yearn for the day it's as fast and cheap as 3.5
2
0
20
@Jason
@ycombinator
@BuildtAi
is an AI tool to help developers search and understand large codebases
1
0
18
First full day in SF, not quite in ‘cerebral valley’ but very cool nonetheless!
2
0
20
I'm trying to write a new system prompt for Buildt and it's crazy how much 'smarter' GPT-4 is than 3.5.
Same prompt for both outputs, 3.5 on the left and 4 on the right. This is with an extensive system prompt and k-shot examples, 3.5 still gets it wrong.
4
0
20
This has made my day, finally I won't need weird ChatML tags to call our codebase search API
We’ve added initial support for ChatGPT plugins — a protocol for developers to build tools for ChatGPT, with safety as a core design principle. Deploying iteratively (starting with a small number of users & developers) to learn from contact with reality:
237
2K
8K
1
0
18
@sdand
Tbf I’m glad they’ve brought this feature out despite being in the current batch, I’ve needed this kind of interface to hook our API up to since they released chat, using ChatML isn’t an elegant solution
2
1
19
It's been a while since I wrote a blog post, so to procrastinate for demo day, which is tomorrow, I wrote this one.
Getting Buildt to work on large codebases, inspired by
@MythBusters
4
1
17
I have experimented with a little hack to get ChatGPT to actually innately **learn** the data I'm fine-tuning it on so it can draw on it when giving answers, a RAG-less approach to retrieval:
2
2
18
The code explanation is also super cool and is really useful if you are new/or rusty on a codebase.
We've also taken the feedback from our initial release and have incorporated those learnings
All of this is live on VS Code extension marketplace:
1
2
17
@houshuang
I've been running it in Node but it should run in the browser yes, I've made it so it has no dependencies at all (I asked for all matrix operations to be done in pure JS which may be slower but makes it more portable, if someone wants to optimize it they can).
I'm just writing…
1
1
17
Well at least a robot found my article interesting 🤣
I built a chatbot to talk to my friends Blog post instead of reading it 🫣->
Okay so
@AlistairPullen
posted a post about the ChatGPT api, but I didn't want to read it all 🤣.
So I built a chat bot in Promptable.js to talk with his article:
5
4
66
2
1
17
Next priority with Buildt: supporting reaaaaaly big codebases
4
0
16
Just had my first instance of an LLM agent going semi rogue on my machine 😅
3
0
15
Quick PSA: If you're having issues with Buildt (again) we're getting swamped with traffic (again), which is very flattering and we're doing our best to fix it!
1
0
16