Steven Heidel
@stevenheidel
Followers
13,469
Following
380
Media
25
Statuses
221
fine-tuning @OpenAI | 🏳️🌈 he/him
San Francisco, CA
Joined October 2014
Don't wanna be here?
Send us removal request.
happy to join microsoft, still not using teams though
380
629
15K
220
252
4K
looking forward to meeting my new intern
24
32
2K
The API team is here. The ChatGPT team is here. The Laundry Buddy team is here. We are all still fully committed to our developers and users.
We are still fully committed to our incredible community of developers and users.
14
67
498
51
36
972
brace yourselves, agi is coming
I'm very excited that today OpenAI adopts its new preparedness framework!
This framework spells out our strategy for measuring and forecasting risks, and our commitments to stop deployment and development if safety mitigations are ever lagging behind.
67
137
975
87
129
916
delve into the latest gpt-4-turbo model:
- major improvements across the board in our evals (especially math)
- dec 2023 knowledge cutoff
Majorly improved GPT-4 Turbo model available now in the API and rolling out in ChatGPT.
451
752
5K
51
40
808
in the future no one should have to do work they don't enjoy
38
53
463
reminder: when you see headlines about other models beating gpt-4 they're likely talking about gpt-4-0314 from last year, which doesn't have the significant improvements we've added to the latest gpt-4-turbo models
44
30
464
❤️
6
11
417
on your last day at openai they give you your shift key back
16
14
421
11
8
383
my twitter replies are a constant battle between team “where’s ilya” and team “when gpt-5”
61
6
384
customers lost = 0
employees lost = 0
❤️ emojis = too many to count
24
9
343
Today at Inflection we are announcing some important updates. A new phase for the company begins now.
Read more here:
54
87
583
12
8
336
the next few months are going to be incredibly exciting
29
16
322
congrats to the
@AnthropicAI
team on the release of Claude 3 today! as i’ve said before the next few months are going to be incredibly exciting 📈
4
16
312
wild that this was generated from a single text prompt
36
22
302
back to work - these models aren’t going to fine-tune themselves!
43
12
297
join us live on this monday at 10am
21
25
368
New! Fine-tuning on functions:
Try it out to reduce the number of prompt tokens spent on function definitions and improve the accuracy of complex function calls. Excited to hear how it works for you!
12
30
242
who’s building AI for taxes? upload all your documents, answer a few questions, output a completed return
75
6
241
again,
16
6
219
legend
12
3
216
💙 i love you all
6
10
210
we are once again so back
🔥Exciting news -- GPT-4-Turbo has just reclaimed the No. 1 spot on the Arena leaderboard again! Woah!
We collect over 8K user votes from diverse domains and observe its strong coding & reasoning capability over others. Hats off to
@OpenAI
for this incredible launch!
To offer
54
208
1K
14
7
209
turns out
@JeffDean
actually wrote all the model weights by hand
#jeffdeanfacts
- seriously though, congrats to everyone at Google and DeepMind on the Gemini release today!
I’m very excited to share our work on Gemini today! Gemini is a family of multimodal models that demonstrate really strong capabilities across the image, audio, video, and text domains. Our most-capable model, Gemini Ultra, advances the state of the art in 30 of 32 benchmarks,
276
3K
13K
6
6
183
it should hopefully be clear at this point that google is not a serious company for developers. i would think twice before building on top of google's cloud services (RIP google domains) or languages/frameworks (Dart, Flutter, Go)
24
11
169
we’re so back
Amazing progress made today. We will come back stronger & more unified than ever:
233
528
8K
6
4
164
It's hard to overstate how much work went into the new OpenAI Assistants API, and how much we owe to the incredible leadership of
@michpokrass
and
@athyuttamre
in getting to today's launch
7
11
162
we’re still so back
“We have unanimously concluded that Sam and Greg are the right leaders for OpenAI.” — Bret Taylor, Chair of the OpenAI Board
The Special Committee of the OpenAI Board has announced the completion of the independent review into November’s events.
589
463
4K
7
6
163
see if you can spot when i switched from manager back to individual contributor
6
1
159
ChatGPT with voice now available to all users! (Breeze is definitely the best voice btw)
ChatGPT with voice is now available to all free users. Download the app on your phone and tap the headphones icon to start a conversation.
Sound on 🔊
2K
3K
18K
13
9
140
openai is nothing without its gross domestic product of several nations combined
7
6
127
reminder to tech recruiters: canada has other universities that aren’t waterloo
19
3
114
lots of talk about openai’s governance yet few understand that the board is ultimately accountable to the bene gesserit
5
5
111
imagine you’re in canada, it’s -40 degrees out, and your furnace breaks. you call the furnace guy, but he can’t make it until tomorrow. so you use
@ChatGPTapp
to upload a few images, then after some back and forth you manage to fix it yourself, avoiding a miserable night.
10
5
94
thanks to everyone who messaged over the past week asking about the hows and whys of fine-tuning - recommend checking out this talk by my colleagues
@john__allard
and
@colintjarvis
3
8
85
you can now stream responses from the assistants API!
Streaming is now available in the Assistants API! You can build real-time experiences with tools like Code Interpreter, retrieval, and function calling.
64
220
1K
10
5
81
enterprise: “we only trust public cloud providers for our AI models”
public cloud providers:
Google Cloud accidentally deleted a company's entire cloud environment (Unisuper, an investment company, which manages $80B). The company had backups in another region, but GCP deleted those too. Luckily, they had yet more backups on another provider.
537
4K
19K
10
7
83
people are using ChatGPT to turn bullet points into emails. then on the other end the receiver is using ChatGPT to summarize the email back to bullet points. this could all be avoided if it were just acceptable to send the bullet points in the first place.
One reason I dislike being sent stuff written by ChatGPT is that it feels like being sent object code instead of source code. The source code was the prompts.
139
99
2K
5
7
79
inpainting is now available in chatgpt!
You can now edit DALL·E images in ChatGPT across web, iOS, and Android.
325
1K
6K
4
2
77
for fun today i tried out the new RealWorldQA benchmark from
@xai
and was able to reproduce the result for gpt-4-turbo (although i got a slightly higher 63.1%). great to see more multimodal evals, keep them coming!
4
1
79
OpenAI 日本
Introducing OpenAI Japan, our first office in Asia, along with a new GPT-4 custom model specifically optimized for 日本語 (the Japanese language).
496
2K
9K
2
1
67
really neat application of AI - recommend checking this out.
@arcinternet
is my default browser on desktop, this may become my new default on mobile.
a little Sunday surprise for you...
meet
@browsercompany
's 2nd product:
🔍Arc Search🔎
it's a default browser for your iPhone
...that BROWSES FOR YOU
the origin story is a bit unusual so I wanted to give you the full backstory:
427
484
4K
3
5
66
interesting (and very readable) paper on improving protection against prompt injection, jailbreaks, and other attacks
Introducing the Instruction Hierarchy, our latest safety research to advance robustness for prompt injections and other ways of tricking LLMs into executing unsafe actions. More details:
176
294
2K
8
3
63
server driven UI + generative AI = ✨. i’ve been looking forward to seeing something like this ever since function calling was released.
2
3
61
found my entire software engineering philosophy summarized in a tweet
It is hard for less experienced developers to appreciate how rarely architecting for future requirements / applications turns out net-positive.
78
702
3K
1
2
57
the price of intelligence is decreasing rapidly
80x cheaper in ~1 year.
In Feb 2023, GPT 3.5 was $20/1M tokens. With batch, GPT 3.5 prompt tokens are now $0.25/1M tokens.
This model class started at $60/1M token (remember davinci). That's 240x in just a couple years.
8
13
168
5
0
57
you can now fine-tune our latest GPT-3.5 model: 0125 (which comes after 0613 and 1106, of course)
We now support fine-tuning for `gpt-3.5-turbo-0125`, our latest GPT-3.5 Turbo model.
Learn more about fine-tuning:
29
99
587
2
1
57
shipped a nice change to
@OpenAI
fine-tuning this week: create jobs without waiting for files to process! we'll start fine-tuning when your files are ready
6
10
56
looks like it’s day 2 at amazon - happy i left when i did
Amazon Prime ads on movies and TV shows will begin in late January
92
26
56
4
1
52
Canada announces $2B investment in AI 🇨🇦
8
4
55
surf instructor in costa rica said to me: “if you look towards your destination you’ll go forward, if you look down you’ll fall down”. that’s my resolution for 2024.
5
1
53
with the new gpt-4-turbo rate limits announced today you can now generate a novel’s worth of tokens every 5 seconds
We’ve just doubled GPT-4 Turbo rate limits across the board and removed all daily limits. Here are the new maximum tokens per minute (TPM) you can now achieve.
109
233
2K
0
3
53
i am excited about chelsea’s excitement
3
0
76
the true measure of your code’s readability is how easy it is for others to change it
5
4
48
recycling reminder! do not put your unused GPUs in the blue bin - instead, place them in a clear plastic bag and send to:
OpenAI
c/o Microsoft Azure
West US 2
thank you for doing your part for the environment 🙏
6
0
44
new from fine-tuning:
- weights and biases integration
- saving multiple checkpoints per training job
- a side by side comparison view in the API playground
- and more!
We're introducing new dashboards, metrics, and integrations in the fine-tuning API to give developers more control, and adding new ways to build custom models with OpenAI.
275
469
3K
4
1
45
the assistants API now supports fine-tuned models, streaming, and 500x the number of files per assistant
Introducing a series of updates to the Assistants API 🧵
With the new file search tool, you can quickly integrate knowledge retrieval, now allowing up to 10,000 files per assistant. It works with our new vector store objects for automated file parsing, chunking, and embedding.
72
323
2K
4
0
45
chatgpt plus is back!
we have re-enabled chatgpt plus subscriptions! 🎄
thanks for your patience while we found more gpus.
803
897
15K
4
5
40
one of the reasons quantum computers are not yet useful is that qubits lose their state too quickly. an exciting result from my former team shows how we can make this state last longer using “dual-rail qubits”
1
3
43
exciting to see quantum twitter and AI twitter coming together tonight
5
0
40
public evals are great, but the best evals for your use case are always going to be the ones you build yourself
Academic benchmarks are losing their potency. Moving forward, there’re 3 types of LLM evaluations that matter:
1. Privately held test set but publicly reported scores, by a trusted 3rd party who doesn’t have their own LLM to promote.
@scale_AI
’s latest GSM1k is a great example.
42
168
891
1
3
40
@charliermarsh
wow - this runs pip install so fast it made me double check that the package was actually installed
1
1
40
Excited to announce that AWS will join the OpenQASM 3.0 Technical Steering Committee, along with IBM, Microsoft, and the University of Innsbruck. Working on quantum assembly languages is an interesting and fun challenge, contributions are welcome!
0
10
39
in the future only robots will be able to prove they are human
4
2
39