Alex Graveley Profile Banner
Alex Graveley Profile
Alex Graveley

@alexgraveley

30,751
Followers
936
Following
99
Media
2,505
Statuses

I’m Alex Graveley, creator of GitHub Copilot, AI Tinkerers, Dropbox Paper, MobileCoin, and Hackpad. Building @ai_minion Hiring

US
Joined July 2022
Don't wanna be here? Send us removal request.
@alexgraveley
Alex Graveley
11 months
My total comp for creating GitHub Copilot, from inception to GA: +20k bonus and a title bump.
456
819
10K
@alexgraveley
Alex Graveley
3 months
If OpenAI is going to continue to eat AI startups sector-by-sector, they should go public ASAP. Building the new economy where only 500 people benefit is a shit future.
167
292
4K
@alexgraveley
Alex Graveley
11 months
Join a startup.
62
107
3K
@alexgraveley
Alex Graveley
11 months
@_elzubeir I’m not upset! Making copilot was an absolute joy. Just putting corpo internals in perspective for others.
12
15
2K
@alexgraveley
Alex Graveley
1 year
Copilot team was only 6 people at time of public release 💪💪💪
45
102
2K
@alexgraveley
Alex Graveley
10 months
School ended.
@frantzfries
Chris Frantz
10 months
Honeymoon over?
Tweet media one
232
83
2K
26
46
1K
@alexgraveley
Alex Graveley
6 months
Deleted pitch slide.
Tweet media one
23
42
1K
@alexgraveley
Alex Graveley
11 months
@moreisdifferent Yes, me and 1-6 others.
16
8
1K
@alexgraveley
Alex Graveley
9 months
It's gets even worse - these copyrighted books are being used to train human writers! A widespread underground network of writing classes and workshops are studying books and repurposing them without royalties or attribution. Shameful.
@dlberes
Damon Beres
9 months
NEW: Meta, Bloomberg, and EleutherAI have trained generative AI on a dataset including upwards of 170,000 pirated books from authors like Stephen King, Zadie Smith, Margaret Atwood. Legality is complex. We have new details and context. tip @Techmeme
35
234
517
25
95
786
@alexgraveley
Alex Graveley
1 year
Semantic search + GPT-3 summary is the todo list app of AI.
29
31
685
@alexgraveley
Alex Graveley
1 year
Glad to see my vision has infected all of Microsoft.
@natfriedman
Nat Friedman
1 year
"Our vision is copilot." -- Microsoft
23
30
979
9
14
674
@alexgraveley
Alex Graveley
7 months
After touching the magic of a working agent, it’s clear that apps and websites are an intermediate phase for humanity.
49
53
616
@alexgraveley
Alex Graveley
1 year
GitHub Copilot exists despite the AI fear on display this week. Fears of productizing Copilot almost halted its creation within OpenAI and led directly to the forking off of Anthropic. Today Copilot helps a lot of people. It would not exist if fear limited access to AI.
29
52
574
@alexgraveley
Alex Graveley
1 year
I run Copilot with autoClosingBrackets and suggestOnTriggerCharacters disabled. Something like 2x the suggestions.
17
29
556
@alexgraveley
Alex Graveley
11 months
@wgussml @moreisdifferent 🙇🏻 correction: + much of OpenAI eng, plus years of their bleeding edge research 🙏
6
4
487
@alexgraveley
Alex Graveley
6 months
Host your own LLM. It’s not that hard. You just need a GPU from the last few years! Finetune and you can shape the exact qualities you want. All it takes is some input-output pairs. Simple! Train your own small models and fill all sorts of niches. They’re super cheap and fast!…
10
40
462
@alexgraveley
Alex Graveley
1 year
I liked $10 and the price sensitivity meter confirmed. $10 means almost everyone can afford Copilot, people rarely cancel, the human feedback data flows, and competitors have slim margins. You’re welcome 🫡
@dannypostmaa
Danny Postma
1 year
I’d pay $499/m for Github Copilot. $10 is a absolute bargain.
174
55
2K
16
21
448
@alexgraveley
Alex Graveley
10 months
I was in Austin the entire time I worked on Copilot. The team was fully remote. SF AI is a vibe, not a requirement.
15
20
416
@alexgraveley
Alex Graveley
1 year
My best guess for LLM in mobile form factor: burn weights into silicon, with last 2 layers in software.
@abacaj
anton
1 year
One day we’ll be able to fit large language models into these little things
Tweet media one
19
10
342
13
13
399
@alexgraveley
Alex Graveley
2 months
🤣
Tweet media one
5
33
386
@alexgraveley
Alex Graveley
7 months
Here’s why your agent doesn’t work: - compounding errors - no trajectory RL - reality doesn’t few shot well - APIs don’t do enough - irrelevant context hurts - subchains destroy nuance And that’s just what we know about.
26
23
378
@alexgraveley
Alex Graveley
1 year
We’re at the beginning of human-computer interaction all over again, at a time before PARC. Significant chance the idea of a “product” or a “UI” doesn’t make sense in a few years.
24
33
364
@alexgraveley
Alex Graveley
5 months
Do you understand now why we called for distributed, decentralized, and unregulated AI ownership?
21
29
354
@alexgraveley
Alex Graveley
8 months
Exited founder high school dropout principal engineer Hiked 2k+ miles Lifted 500 lbs Fell in love 40+ countries visited Cancer survivor Hitchhiked across SE Asia Been rich, been poor Seeking does not lead to finding. It’s always an inside job 🙏🏻
18
10
328
@alexgraveley
Alex Graveley
5 months
Treating engineers as commodity is symptom of promoting to mgmt track too early. Predictive in every zombie tech co. The answer is small teams, high autonomy, and firing fast.
@nabeelqu
Nabeel S. Qureshi
5 months
Ouch. Rare to see someone being this publicly frank about problems at a large tech company (here Google):
Tweet media one
43
114
2K
4
18
326
@alexgraveley
Alex Graveley
6 months
People said this as a reason we shouldn’t release GitHub Copilot. It’s been a few years - show me the harm?
@ChrSzegedy
Christian Szegedy
6 months
I think Yann might underestimate the potential of AI if people have API access to strong generative AI. LLMs are capable of generating code which could be executed *automatically* by *anyone* without any human *oversight*, also in a loop and open-endedly. This is very hard to…
43
61
483
8
17
318
@alexgraveley
Alex Graveley
8 months
gpt-3.5-turbo-instruct brings back logprobs 🥳🙏🏻 Finally chance of making a real product again.
12
19
299
@alexgraveley
Alex Graveley
6 months
Agents are the prize and no one has cracked it. Don’t listen to the hype.
21
23
277
@alexgraveley
Alex Graveley
5 months
Q*: Trading inference time MCTS for model capacity. Meaning you can spend 1000x time picking the next token, to approximate a model 1000x the size. Which you then distill down to a today-sized model. Implications for sample efficient self-play. I wish I understood what any of…
10
21
278
@alexgraveley
Alex Graveley
11 months
Copilot was made using startup principles, with a tiny team, in under a year, inside very dysfunctional GitHub/MSFT org. Imagine how much good could be unleashed within existing orgs if they put trust in individuals instead of hierarchy.
@paulg
Paul Graham
11 months
Talked to a programmer today who said AI coding tools made him about 10x more productive. Though 10 seems like a round number, this was an attempt at a precise estimate.
387
150
2K
3
26
251
@alexgraveley
Alex Graveley
1 year
The 🐐 speaks on prompt engineering
4
33
252
@alexgraveley
Alex Graveley
1 year
Will go a little further than @simonw . You can run a 65bn param model on your Mac. In a few weeks there will be serviceable Copilot and chatGPT you can run yourself. We are now in the awesome timeline.
@simonw
Simon Willison
1 year
OK, I'm calling it: Large language models are having their Stable Diffusion moment right now
45
426
2K
17
19
249
@alexgraveley
Alex Graveley
1 year
Consider using YAML instead of JSON for user-visible LLM output, since you can't stream-parse JSON easily.
30
11
247
@alexgraveley
Alex Graveley
1 year
@jobergum You answered your own question :)
0
0
239
@alexgraveley
Alex Graveley
5 months
Everything is going to run on mixtral soon. Perf : cost : speed is bonkers.
12
12
237
@alexgraveley
Alex Graveley
3 months
@ohshitdatboee Microsoft IPOd in 1986, a year after releasing Windows.
2
0
221
@alexgraveley
Alex Graveley
1 year
UL2 vs. LLaMA
Tweet media one
Tweet media two
15
11
208
@alexgraveley
Alex Graveley
1 year
1 dedicated madman makes a project, 2 makes a startup, 3 a unicorn, 5 a google. Getting 5 inside any company is insanely rare.
@collision
John Collison
2 years
As you become an adult, you realize that things around you weren't just always there; people made them happen. But only recently have I started to internalize how much tenacity *everything* requires. That hotel, that park, that railway. The world is a museum of passion projects.
670
10K
53K
3
15
196
@alexgraveley
Alex Graveley
5 months
Now that my X algo is fully AI saturated, X is as addictive as tiktok 😬
18
3
199
@alexgraveley
Alex Graveley
1 year
Met 3 transformers paper authors so far on this trip to SF. Feeling awestruck and insanely grateful 🙏🏻
2
2
186
@alexgraveley
Alex Graveley
1 year
👀 “32x larger context lengths” “500B+ parameter models… low-batch-size latency of 29ms” 🎁
@_akhaliq
AK
1 year
Efficiently Scaling Transformer Inference abs:
Tweet media one
0
26
125
4
11
186
@alexgraveley
Alex Graveley
10 months
Story time. In 2016, I deployed my first AI system. I had wanted to work on AI Personal Assistants, and joined a text-based PA company. We trained a small seq2seq model on all the chats, and for the PAs it would suggest responses. (Kind of like Copilot grey text). It worked…
@alexgraveley
Alex Graveley
10 months
Part of the motivation with meetups is to remove the class divide between AI tinkerers, practitioners, researchers. Also break the Bay Area centrism. Make place to share openly and build high-trust community. Mixed results so far, looking for new ideas!
8
4
51
8
14
186
@alexgraveley
Alex Graveley
6 months
I'm looking for a hacker. Offer is a lot of equity and work with bleeding edge AI. DM me.
22
10
181
@alexgraveley
Alex Graveley
5 months
No one is asking why an App Store is so important to OpenAI - it’s required for their version of AGI. It will teach the model when and how to use external APIs, which they can weave together as demonstrated with the code-interpreter/dall-e APIs.
24
10
184
@alexgraveley
Alex Graveley
1 year
Stumbled on my first ever demo of Copilot "ghost text" from Jan 2021.
11
9
179
@alexgraveley
Alex Graveley
7 months
Come stay in my scenic SF apartment Airbnb, 2x8 A100 80gb nodes included, great views and 5 min walk to the Mission. Relax while you train models in urban luxury. - $1500/night
8
5
175
@alexgraveley
Alex Graveley
1 year
Using @karpathy theory of "time to think", asking for a list & then a condensing the list (in the same completion) yields a better result.
Tweet media one
Tweet media two
5
13
168
@alexgraveley
Alex Graveley
1 year
LLaMA RLHF demo at AI Tinkerers SF rn
Tweet media one
6
7
169
@alexgraveley
Alex Graveley
1 year
If you care about AI Safety, fire up LLaMA at home. Start poking around at activations. Make open analysis tools and test suites. Where do slurs, apocalypse ideas, biases live in the network? How to remove them without hurting perf? YOU can help make AI safe.
10
13
167
@alexgraveley
Alex Graveley
7 months
Google is hiring smart people to work on agents because it's the one thing that can disrupt their monopoly on the web. A fundamentally defensive strategy.
@bonniesjli
Bonnie Li
7 months
Excited to share I've joined @GoogleDeepMind ! Looking forward to working with amazing researchers and building generally intelligent agents! 🚀
42
8
421
4
8
169
@alexgraveley
Alex Graveley
1 year
Congrats to Copilot team! My baby is all grown up! 🤩
@satyanadella
Satya Nadella
1 year
GitHub Copilot for Business is now available, as we bring the world’s first at-scale AI developer tool to any organization.
119
417
3K
7
0
164
@alexgraveley
Alex Graveley
1 year
Generate example task Break down into steps Loop ( Generate code for step Run code Expose error ) Save working code as kshot example If this is exciting to you, we're hiring @ai_minion 🤖
@ai_minion
Minion AI
1 year
Hello, World!
14
13
133
8
10
159
@alexgraveley
Alex Graveley
11 months
Looking for a web-focused full stack senior engineer to join @ai_minion in SF. Willing to salary match. Join us to work at the bleeding edge of AI: prompting, fine-tuning, synthetic data, codegen, planning, reasoning, and memory for embodied agents. jobs+eng @minion .ai or DM.
13
21
160
@alexgraveley
Alex Graveley
1 year
Would gladly pay for premium Google search that downranks SEO sites.
10
3
154
@alexgraveley
Alex Graveley
10 months
I'll pay $2k to anyone who fixes this, e.g. merged PR to convert pgvector to hnswlib. Anyone else want to pile on?
@NirantK
Nirant
10 months
Why you should never use pgvector (e.g. @supabase Vector Store) for production: 😮 pgvector is 20x slower than a decent vector DB (e.g. @qdrant_engine ) 🤯 And it's a full 18% worse in finding relevant docs for you And this can happen at as little as 10K documents when chunked!
Tweet media one
33
30
217
21
14
157
@alexgraveley
Alex Graveley
1 year
Text is not the universal interface. 🧵
Tweet media one
3
7
156
@alexgraveley
Alex Graveley
1 year
All code is bad. Some is just currently maintained.
7
13
151
@alexgraveley
Alex Graveley
1 year
Conspiracy mind says Elon’s plan all along is to make a new phone. Phone + twitter + starlink. Owning the entire stack of human communication.
20
4
149
@alexgraveley
Alex Graveley
10 months
Llama 2 flips the gameboard. I’m excited! Also grateful to not have raised 100m+ to make a foundation model.
10
1
151
@alexgraveley
Alex Graveley
5 months
Has anyone looked into “attention expansion” where you’d replace a large section of prompt with an ellipsis token, and if attention on ellipsis is high you expand with original content? Ideally automatically without having to repeat inference.
13
6
151
@alexgraveley
Alex Graveley
11 months
4
0
143
@alexgraveley
Alex Graveley
6 months
One way to think of the Open in OpenAI is that they are teaching a lot of devs about ML/LLM through the pacing of their releases. Prompting, context length, attention, RAG, finetuning, param count, datamix, tok/s... all these concepts unknown outside academia a year ago.
12
7
146
@alexgraveley
Alex Graveley
10 months
Something different about AI projects: it’s best to iterate by starting over from scratch when you find a better abstraction. Just cannibalize the old stuff and delete it. Inaccuracy compounds with the wrong metaphor & slows dev down.
6
10
145
@alexgraveley
Alex Graveley
9 months
The people tinkering with small LLMs are undervalued by industry rn: AI apps become ensembles of small, fast dedicated models in support of big honking models.
9
9
140
@alexgraveley
Alex Graveley
1 year
I didn’t accept my Twitter offer letter in 2011. After a 2 week try out I was the only one in the office late - seemed no one wanted to work hard.
5
2
140
@alexgraveley
Alex Graveley
1 year
I’ve never started a project where I didn’t feel lost and confused, incompetent and stupid, questioning the premise 1000 times.
7
11
139
@alexgraveley
Alex Graveley
1 year
It’s not “Copilot for X” if you have to remember to break out of your usual flow to use it.
@alexgraveley
Alex Graveley
1 year
@gdb Copilot is automatic, this is user-directed. Big difference.
4
0
24
6
5
139
@alexgraveley
Alex Graveley
1 year
Wait until you see what I’m up to next 😉
13
1
137
@alexgraveley
Alex Graveley
2 months
People dunking on Google for messing up historic images should spend a day imagining how to solve the problem. Draw out a little decision tree to see how hard it is. Fact vs fiction vs diversity vs helpful vs problematic isn’t easy to calibrate.
31
8
139
@alexgraveley
Alex Graveley
10 months
Best possible solution: upload all public tweets in parquet file daily. Keep the login wall up permanently. Minimal ops impact, zero incentive to scrape.
@elonmusk
Elon Musk
10 months
@TimSweeneyEpic Several hundred organizations (maybe more) were scraping Twitter data extremely aggressively, to the point where it was affecting the real user experience. What should we do to stop that? I’m open to ideas.
5K
2K
26K
12
6
132
@alexgraveley
Alex Graveley
11 months
As best anyone can tell, advanced reasoning in LLMs comes from learning code. Are humans the same?
33
7
129
@alexgraveley
Alex Graveley
1 year
Looking to hire a full-stack dev and a designer who want to crank on a new AI product idea. DMs open.
12
18
130
@alexgraveley
Alex Graveley
11 months
@JThomasBurgess No, Shuyin is awesome.
1
1
129
@alexgraveley
Alex Graveley
1 year
Has anyone hooked up GPT-3 to XGBoost yet?
10
5
119
@alexgraveley
Alex Graveley
6 months
Emerging combination of LLM + RL + codegen for (agent || robot) is an interesting unification I didn’t expect. Seems like many/most AI systems in the future will be some variant of this?
2
5
122
@alexgraveley
Alex Graveley
7 months
We do these things not because they are easy, but because we thought they would be easy.
0
9
120
@alexgraveley
Alex Graveley
8 months
I think there’s a serious new inequality introduced by Copilot et al: adding code is much easier than refactoring. Previously they were comparable. Top programmer now means keeping the system in your head, not being lazy, finding clean abstractions.
14
5
120
@alexgraveley
Alex Graveley
5 months
Has anyone tried using a LoRA per user instead of a long term memory system?
24
6
120
@alexgraveley
Alex Graveley
7 months
Super useful GPT self-calibration prompt examples at the end of
Tweet media one
1
13
117
@alexgraveley
Alex Graveley
1 year
Looking for half-baked blog post feedback. Is this already obvious to everyone?
Tweet media one
37
9
113
@alexgraveley
Alex Graveley
1 year
I’ve been using this trick for years. It works. “Before stepping away, leave the code in a state where it is Obviously Broken, but Easy to Fix.”
4
18
113
@alexgraveley
Alex Graveley
6 months
+5% on evals this weekend. Finetuning on all data, then again on high quality. TBD is boost from DPO - gives +8% on small model. Proved we are data limited. Scaling up self-play next week.
6
2
113
@alexgraveley
Alex Graveley
1 year
Lol this blew up. Copilot not possible without the geniuses at OpenAI and the principled VSCode editor crew. Sorry if I forgot anyone! 😅
3
1
108
@alexgraveley
Alex Graveley
10 months
There’s room for an agent foundation model.
11
6
107
@alexgraveley
Alex Graveley
4 months
2024 resolution: ruthlessly cut out anyone perpetuating “America is over” meme. 🇺🇸
15
12
108
@alexgraveley
Alex Graveley
6 months
OpenAI is 🐐
4
2
102
@alexgraveley
Alex Graveley
11 months
> you require a $500k per month commitment to use Azure GPT4-32k? < Yes, that would be the approximate amount. Azure OAI NGMI.
21
4
100
@alexgraveley
Alex Graveley
1 year
One would hope for front page of WSJ to mention the extreme productivity gains, but I’ll take it anyway! 🥰
Tweet media one
9
8
96
@alexgraveley
Alex Graveley
1 year
New canned response when people ask me to help them with their product: “Yeah I don’t care about any of that marketing crap. Tell me about the core technical problems, how you’re breaking them down to solve them, your key metrics, and where your baseline performance is today.”
5
4
93
@alexgraveley
Alex Graveley
9 months
When I met @pmarca he asked me about how multi-agent systems would collaborate in the future. My response: I’m just trying to order pizza online reliably 🤣
7
1
97
@alexgraveley
Alex Graveley
6 months
@markchen90 Consider an outsider's perspective: OpenAI has progressively published fewer details on it's models. It was non-profit, now for-profit. It routinely ships features that its customers sell. Now it it contributing to hindering others from replicating.
2
7
93
@alexgraveley
Alex Graveley
3 months
Off X for a bit. We hit our target accuracy of 85% on a top 10 website. Time to crank.
9
1
94
@alexgraveley
Alex Graveley
11 months
Language skills are now humanity’s bottleneck - LLM interfaces will change this drastically in a generation.
1
10
89
@alexgraveley
Alex Graveley
4 months
Worked for us on llama2 as well!
@sdand
surya
4 months
i implemented Self-Extend on Mistral 7B last night it extends context length without fine-tuning from 8k to 16k and more using a new bi-level attention technique based off group attention code: paper:
Tweet media one
11
64
633
3
9
92
@alexgraveley
Alex Graveley
8 months
Simple LLM technique that helps a lot (but you might not be using): add a constraint checker to ensure valid generation. On violation, inject what was generated and the rule violation, and regenerate.
5
4
92
@alexgraveley
Alex Graveley
1 year
Excited to see everyone at The Commons in SF tonight starting at 6 🤗 Based on the austin/seattle meetups we expected ~20, and rented a space to support 80 people. We currently have 450 people who have RSVPd 🤯 I’ll be at door to personally apologize to people we can’t let in…
@alexgraveley
Alex Graveley
1 year
👋🤖 AI Tinkerers SF meetup happening 2/22 - see you there!
13
5
77
5
1
91
@alexgraveley
Alex Graveley
8 months
That feeling when everything aligns and you’re 100% certain you’re holding the future. Cosmic? Enlightened? Felt it with ghost text prototype for Copilot. Felt it last week with @ai_minion . Cannot wait to get this in people’s hands.
8
2
89
@alexgraveley
Alex Graveley
1 year
Tweet media one
@goodside
Riley Goodside
1 year
I stand with Rahul Ligma.
Tweet media one
11
9
411
2
0
88
@alexgraveley
Alex Graveley
6 months
Sama biggest rock star since Jobs 🫡
3
2
88