Boris Power
@BorisMPower
Followers
34K
Following
10K
Media
176
Statuses
2K
Head of Applied Research @OpenAI
San Francisco, CA
Joined July 2017
This release is much closer to the original GPT-3 release than the ChatGPT release. A new paradigm, which some will find incredibly valuable for things even we at OpenAI can’t predict. But it’s not a mass product that just works and unlocks new value for everyone effortlessly.
25
69
1K
And we love you ❤️.
i love the openai team so much.
24
55
3K
Mistakes happen. Thank you Ilya, this is inspiring.
I deeply regret my participation in the board's actions. I never intended to harm OpenAI. I love everything we've built together and I will do everything I can to reunite the company.
50
69
2K
This all happened after midnight, during a company wide holiday. Support was always near unanimous.
38
166
2K
We just released new embeddings at a ridiculously low price!. For $40 one can embed all Joe Rogan podcast's transcripts (100k pages of text!) - who will go viral by building a semantic search for it?. This is better, and 150x cheaper than previous models!
50
188
2K
Some more napkin math - size of the Internet is ~10^11 pages of text*, this would cost (only?) $50M to embed. Who wants to take on Google?.
We just released new embeddings at a ridiculously low price!. For $40 one can embed all Joe Rogan podcast's transcripts (100k pages of text!) - who will go viral by building a semantic search for it?. This is better, and 150x cheaper than previous models!
46
215
2K
We just open sourced an incredible speech recognition model. It's generally better than me, and works incredibly robustly across background noise, accents and language mixtures, which is really exciting! Let's use it everywhere!
24
224
1K
Recreating OpenAI reporting structure / social graph based on ♥️s
26
33
1K
I’m on a research visa too that i will lose if i resign. These are details - onwards with the mission! 🚀.
I am on an H-1B, in the process of getting my green card and relocating my family to the US. Me and many other colleagues in a similar situation have signed this letter. I do not know what will happen next, but I am confident we will be taken care of. the board should resign.
37
59
1K
@Dr_Singularity I’m sorry we failed you and thanks for the patience - hopefully we rectify this soon and make the subscription way more valuable.
113
16
695
The real heroes are the unsung heroes who kept up the website, security and infrastructure through all this! ❤️.
19
35
613
“Majorly improved” 😉.
Majorly improved GPT-4 Turbo model available now in the API and rolling out in ChatGPT.
22
9
548
Completely wild that the same capability underlying model (and fine tuning) was available through the api for 8 months prior and nobody jumped an opportunity to do this, so we had to build it ourselves. What’s the next UX innovation, waiting to happen in this space?.
This is insane. This is what I’ve been alluding to for months now. This is an epochal transformative technology that will soon touch - and radically transform - ALL knowledge work. If most of your work involves sitting in front of a computer, you will be disrupted very, very
44
56
524
We increased rate limits for o1-preview and o1 mini for tier 5. We'll be increasing the limits over the coming weeks, and introducing access to more tiers. 🚀. o1-preview -> 100 requests per minute.o1-mini -> 250 requests per minute.
19
26
526
Every paid ChatGPT user should now have access to OpenAI o1-preview!.
39
15
491
Let us know how we can improve the experience of talking to ChatGPT using voice!.
rollout completed early, amazing work by the team!. (except for jurisdictions that require additional external review).
313
11
490
@vkhosla @realDonaldTrump You read the wrong stat. Unemployment rate in Argentina is around 7%. The article refers to “poverty rate”, which is a very different measure.
1
5
489
Taking what did you get done this week to the extreme. Would it be a new normal for efficient companies to release something new each day?.
12 days. 12 livestreams. A bunch of new things, big and small. 12 Days of OpenAI starts tomorrow.
34
12
458
Go try out o1-mini and don’t be put off by the name - it’s better than o1-preview for some tasks !. It’s really really good and fast.
29
16
438
Today we released #dalle 2 - a model which can generate incredibly impressive images based on a textual description!. "A cobra, surfing on a big wave". (Feel free to drop suggestions in the thread - I'll generate and share if they are fun!)
128
47
406
I got fooled by the Reflection 70B announcement. tl;dr - the model performs very badly.
A story about fraud in the AI research community:. On September 5th, Matt Shumer, CEO of OthersideAI, announces to the world that they've made a breakthrough, allowing them to train a mid-size model to top-tier levels of performance. This is huge. If it's real. It isn't.
18
24
413
Lots of tweet from almost everyone involved, clarifying the situation and position. But i haven’t heard from three board members at the center of this.
22
28
404
Solving intelligence is not enough. It seems like that as our civilization was bottleneck by intelligence since the Industrial revolution. Intelligence will soon become near free - what will be the next bottlenecks?.
202
23
391
It’s a new state of the art - please try it!.
8
17
371
I’d love to collect some feedback. Would you mind responding to this tweet with examples that you were truly IMPRESSED by?.
92
18
377
Two ways to get GPT-4 access 🤖. 1. Sign up to the waitlist (it’ll take a few weeks)🤞.2. Contribute an open source eval (instant access if the PR is approved!) 🎉.
16
40
366
I can’t even imagine what your next thing could possibly be but i trust it’ll make OpenAI just a footnote!.
i loved my time at openai. it was transformative for me personally, and hopefully the world a little bit. most of all i loved working with such talented people. will have more to say about what’s next later. 🫡.
5
11
358
Robotics is still very far from being able to do anything close to this. Excellent goal oriented long term planning, agility and replanning when necessary.
11
36
362
@paulg Keep in mind that 3 is not even a majority of the starting 6 from when this started less than a week ago.
7
8
354
Transformers and diffusion models are way too parameter heavy to be practical. A simple temporal sequence Boltzmann machine can simulate the videos of bouncing balls 100,000x more efficiently than Sora. See this gem of a sleeper paper for more details!.
14
30
355
First time proper RL over the space of language! This brings up memories of early days of computers getting better at Go via self play.
We're releasing a preview of OpenAI o1—a new series of AI models designed to spend more time thinking before they respond. These models can reason through complex tasks and solve harder problems than previous models in science, coding, and math.
4
14
351
Thank you everyone for the support that you have expressed to me privately. It is very appreciated! 💜 . I never knew so many customers, friends, past colleagues, and strangers are watching and cheering for us. Looking forward continuing these amazing relationships wherever i end.
20
10
347
Five in a row this week! Have fun!.
Day 3: Something you’ve been waiting for.
16
8
347
Amazing investigative work! 👏 😆.
13
16
337
@Benioff @salesforce @salesforcejobs Lol, like it was ever about compensation. We got >95% in less than 24 hours, and the compensation never crossed my mind!. Can you give us 700 of our amazing colleagues, then we may budge 🚀.
8
7
320
Llama-2-70b costs $59 per million tokens.GPT-3.5 Turbo costs $2 per million tokens.
23
17
323
Whoever has ⏸️, ⏹️, ⏩, or ▶️ in their Twitter name is overly radicalized and is difficult to take seriously.
80
12
316
Good luck - humanity will benefit immeasurably if you succeed!.
Superintelligence is within reach. Building safe superintelligence (SSI) is the most important technical problem of our time. We've started the world’s first straight-shot SSI lab, with one goal and one product: a safe superintelligence. It’s called Safe Superintelligence.
7
6
319
For everyone who wants to use o1 effectively - the actual gold is the chains of thought in the blogpost. The amount of detail and rigor helps you appreciate how the algorithm works, and what it may be capable of!.
16
15
323
Congrats on the "low key research preview" of computer use! ;).
Introducing an upgraded Claude 3.5 Sonnet, and a new model, Claude 3.5 Haiku. We’re also introducing a new capability in beta: computer use. Developers can now direct Claude to use computers the way people do—by looking at a screen, moving a cursor, clicking, and typing text.
9
13
315
@apples_jimmy First time an LLM provider increases prices for a released model. Interesting and unusual.
15
4
306
It was incredible to see the open mindedness of OpenAI employees and leadership, in stark context to ideological dogmatism of the board. We listened, tried to follow along and understand. But it was always just a power grab. Well, only one lesson can be learnt from this. Onwards.
10
19
302
This demo from Khan academy is a glimpse into the future of education!. Any one out of the tools featured in this video is 🤯 but all of them together honestly make me want to learn school subjects all over again in this fun exploratory way with GPT-4!.
7
53
280
@rishabh16_ Our models are available to anyone. Who will build the next gen search tools and commercialise them effectively?.
17
19
273
OpenAI was able to strongly relax our policies by:.1. aligning our models to be great at following instructions 2. releasing a moderation endpoint, to help developers Almost anything net-good is allowed now!
The reason we’re seeing a resurgence in the GPT-3 apps is because OpenAI took the training wheels off. No more restrictions around long-form writing or social posting using their api *and* you don’t need to be manually approved. You can pretty much do anything now.
10
34
250
Long awaited, true delightful voice experience is finally available!.
Advanced Voice is rolling out to all Plus and Team users in the ChatGPT app over the course of the week. While you’ve been patiently waiting, we’ve added Custom Instructions, Memory, five new voices, and improved accents. It can also say “Sorry I’m late” in over 50 languages.
22
4
251
Model distillation easily enabled on our platform! Distill an o1-preview model into 4o-mini on your api traffic!. And an evals product to make it easier to understand the tradeoffs!.
12
8
252
An insider’s view attempting to explain to untrained mathematicians how hard FrontierMath dataset is. Average human performance given unlimited time is approximately 0%.
1/11 I’m genuinely impressed by OpenAI’s 25.2% Pass@1 performance on FrontierMath—this marks a major leap from prior results and arrives about a year ahead of my median expectations.
13
25
248
It only gets better from here!.
🤯🤯🤯I'm shocked by the results from OpenAI's o1 model on THIS YEAR's Korean SAT exam. It got only *ONE* question wrong, placing it within the Top 4% of students. This exam was crafted by professors who were locked up in a hotel for a month, making it an unseen test set for all
12
17
236
Easiest way to get 32k context GPT-4 model access is to submit a good open source Eval that’s accepted!. Evals that creatively utilize long context will be prioritized for review.
9
22
226
Can’t wait to see how the new updates are received externally! I’m excited!.
We’ll be streaming live on at 10AM PT Monday, May 13 to demo some ChatGPT and GPT-4 updates.
10
3
221
the most beautiful scaling plot in the past 5 years.
@OpenAI o1 is trained with RL to “think” before responding via a private chain of thought. The longer it thinks, the better it does on reasoning tasks. This opens up a new dimension for scaling. We’re no longer bottlenecked by pretraining. We can now scale inference compute too.
3
14
226
The fastest growing open source repo in OpenAI history, almost entirely built by Isabella in a few weeks!.
🌟 Our ChatGPT Retrieval Plugin is #1 trending on GitHub! Improve your ChatGPT experience by accessing relevant information from your personal or organizational documents using OpenAI's embeddings models and vector databases. Here's a quick overview:.
2
16
222
Join us in an hour!.
We’ll be streaming live on at 10AM PT Monday, May 13 to demo some ChatGPT and GPT-4 updates.
7
8
216
@aidan_mclau The reason chess engines can do it is they are essentially a hardcoded agent that keeps applying simple functions. Similar to a human - given a pen and paper we can think for a lot longer about one problem. The exciting thing about o1 is that it’s reliable enough for agents.
10
10
217
I feel sorry for the original voice actor who has to go through and withstand this drama - it doesn’t seem fair to have your own voice questioned like this. 💜.
The Sky voice was available since September of last year. It has been used by millions of people in ChatGPT without an issue. The controversy began when it all of a sudden had more personality and inflection because of GPT-4o. It could have sounded like a Kardashian and people.
14
16
212
A very clear prediction of what the future of intelligence and the effect on the world will be by Sam Altman. "the future is going to be so bright that no one can do it justice by trying to write about it now". The Intelligence Age is upon us!.
12
13
208
@alexandr_wang I applaud this effort. But don’t call it last, otherwise we’ll get to FINAL_FINAL_v5_latest.
2
1
205
It would be great if someone figured out how to make super hard evals economically a viable business. Every lab would pay 💰.
3/10 We evaluated six leading models, including Claude 3.5 Sonnet, GPT-4o, and Gemini 1.5 Pro. Even with extended thinking time (10,000 tokens), Python access, and the ability to run experiments, success rates remained below 2%—compared to over 90% on traditional benchmarks.
18
7
196
Evals are difficult to come by - here’s a good one we’ve used internally that’s not yet saturated!.
Factuality is one of the biggest open problems in the deployment of artificial intelligence. We are open-sourcing a new benchmark called SimpleQA that measures the factuality of language models.
4
3
196
@pwang_szn @OpenAI @LeagueOfLegends Really cool!. But the commentary isn’t very good. It’s a bit like chess commentary from gpt4. Lots of obvious statements and a few hallucinations. Missing the key aspects - two towers dropped on opposite sides and top lane had to vacate the tower due to dive potential.
5
0
188
@Dr_Singularity Are you sure? Maybe the world will be hyper deflationary, making your current money orders of magnitude more valuable in that future world.
17
4
193
For example i was on an 11 hour flight and did my best to understand the situation virtually.
7
3
185
Exactly. And is getting better at an impressive speed.
O1 is just godly good. I uploaded my blood test results and discussed with it all possible therapies. It’s literally the best doctor I ever had. Not only are the diagnoses and therapies on point but I also learn so much about my body and myself as o1 explains everything in.
8
6
188
Wow, nice work Aravind! How will this be different / better than the interactive map on .
Yes, it is! There's only one AI that can do this for Tuesday!
12
3
185
A great way to round up the 12 announcements!.
Day 12: Early evals for OpenAI o3 (yes, we skipped a number).
9
3
183
Finally integrated search within a beloved interface!.
🌐 Introducing ChatGPT search 🌐. ChatGPT can now search the web in a much better way than before so you get fast, timely answers with links to relevant web sources.
12
6
179
We’re in great hands - congratulations, this is very well deserved! 🚀.
2
2
178
I love the difference between o1-mini and o1-preview!.
Prompt: 080525 081523 011805 251521. o1-mini: Pass.o1-preview: Pass.GPT-4o: Fail
8
4
177
Phi-3 can run on a phone, and is very competitive with gpt-3.5!.
Phi-3 Technical Report. Microsoft presents a new 3.8B parameter language model called phi-3-mini. It's trained on 3.3 trillion tokens and is reported to rival Mixtral 8x7B and GPT-3.5. Has a default context length of 4K but also includes a version that is extended to 128K
17
18
175
@fchollet Regardless if we solved it or not - I’d love to figure out a way to incentivize the creation of much harder evals that truly stand unsolved for more than a few months, and are ideally economically valuable.
18
4
171
Can’t wait to see what you do when the voice + video streaming is implemented in our API !.
We’ve just added GPT-4o to the Ai Bus — expect your Pin to be even better at understanding your requests. This is live on all Ai Pins — no action needed. The architecture of CosmOS enables us to quickly adopt the latest and greatest AI technologies.
10
6
171
SearchGPT is out today! Excited to hear your feedback. I’m very impressed with the speed and ability to drill deeper quickly - something the old search doesn’t serve well.
22
7
169
Remember, this is the worst the technology will ever be ;).
I gave ChatGPT a screenshot of a SaaS dashboard and it wrote the code for it. This is the future.
3
11
169
Incredible, I was hoping Andrej releases something like this, but i really didn’t think it would be as comprehensive as this!. Making latest innovations as accessible as this is just 🔥!.
📽️ New 4 hour (lol) video lecture on YouTube:."Let’s reproduce GPT-2 (124M)". The video ended up so long because it is. comprehensive: we start with empty file and end up with a GPT-2 (124M) model:.- first we build the GPT-2 network .- then we optimize
4
8
170