Nikunj Handa
@nikunjhanda
Followers
2K
Following
7K
Media
41
Statuses
674
1. Vision + function calling.2. Dec 2023 cutoff.3. *Huge* improvements across the board in our evals (particularly in math!).
GPT-4 Turbo with Vision is now generally available in the API. Vision requests can now also use JSON mode and function calling. Below are some great ways developers are building with vision. Drop yours in a reply 🧵.
9
7
200
Excited to share what I've been working on at OpenAI! Lots of improvements to make it a lot simpler for anyone to build powerful AI applications with out-of-the-box retrievals, code interpreter, and function calling. Would love your direct feedback - nikunj@openai.com.
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.
13
11
149
Really excited to launch speculative decoding in the OpenAI API! 🔮⚡️. For tasks involving rewrites, if you already have some idea of what a response would be (e.g. the original draft), pass it through `predictions` and you'll see a significant speed-up.
Introducing Predicted Outputs—dramatically decrease latency for gpt-4o and gpt-4o-mini by providing a reference string. Speed up:.- Updating a blog post in a doc.- Iterating on prior responses.- Rewriting code in an existing file, like @exponent_run here:
2
3
119
So excited to be releasing o1 in the API 🍓 . It's been amazing to see everything our early testers have already achieved with this model in coding, legal, healthcare, and education. We'll be increasing rate limits, expanding availability, and adding new features in the coming.
OpenAI o1-preview and o1-mini are rolling out today in the API for developers on tier 5. o1-preview has strong reasoning capabilities and broad world knowledge. o1-mini is faster, 80% cheaper, and competitive with o1-preview at coding tasks. More in
4
6
111
This has been a labor of love! Excited to see everything you build with 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.
5
0
66
Loving the reception on o1 so far—we hope this allows more of you to go to production with the model!. Keep the feedback coming (DMs are open)!.
Just 5x'd rate limits again: . o1-preview: 500 requests per minute .o1-mini: 1000 requests per minute.
1
0
58
Incredibly proud of all the ships the API team worked on for devday 2024—from realtime api, to *automated* prompt caching, to distillation, to vision fine-tuning. Little big feature: easily generate prompts, function definitions, and structured outputs in the playground (h/t.
Today at DevDay SF, we’re launching a bunch of new capabilities to the OpenAI platform:
4
0
54
Super excited to be bringing DevDay home to Singapore 🇸🇬 — see you there on Nov 21!.
We’re taking OpenAI DevDay on the road! Join us this fall in San Francisco, London, or Singapore for hands-on sessions, demos, and best practices. Meet our engineers and see how developers around the world are building with OpenAI.
3
1
28
@simonw We're still working on getting o1 live in the API — lots of new features to add so stay tuned!.
1
3
28
My only wish for Elon's Twitter is for it to stop flagging all the @ChessvisionAI tweets as 'offensive'. It has been weeks!
4
0
19
@itaybachman @OpenAIDevs We hope to expand access to more usage tiers over the coming weeks!.
1
1
22
*hundreds of thousands* 🤯.
Since we started offering free fine-tuning, developers have fine-tuned hundreds of thousands of OpenAI models. Free fine-tuning was supposed to end yesterday, but we've decided to extend it through October 31. 🔧👻. 4o: 1M free training tokens a day .4o mini: 2M free training.
2
0
22
We're adding more knobs to our hosted RAG tool -- you can now customize chunking and the number of chunks retrieved. Which other knobs would you like to see?.
You can now customize how your files are chunked and change the number of chunks returned in the Assistants API’s File Search tool.
5
0
21
Vision in now available in the Assistants API! h/t @erinkav . What else would you like to see in the Assistants API?.
The Assistants API now supports vision. You can create messages with image URLs or uploaded files, and your assistant will use the visuals as part of its context for the conversation.
5
0
21
This is huge -- something much better and faster than GPT-4T is also 50% cheaper. Available via the Assistants API today -- can't see what you build with it!.
in the API, GPT-4o is half the price AND twice as fast as GPT-4-turbo. and 5x rate limits.
4
0
18
@altryne @OpenAIDevs We will definitely add support for streaming and are considering adding support for reasoning progress. Thanks for your feedback!.
0
0
18
Excited to have @chughesjohnson visit @StripeSingapore and #tiasg2018 tomorrow! Every conversation I've had with Claire has always been filled with invaluable advice and feedback—highly recommend all to attend if you're in town! Details:
0
0
17
0
1
15
To engineers living in / interested in moving to SG: cant recommend @StripeSingapore enough. This is an amazing opportunity with outsized ownership & growth potential — we’re just getting started. Reach out directly if you’re keen on learning more (nikunj@stripe.com).
0
2
13
Great time speaking at the #AWSstartupday today!
0
1
13
@ai_for_success @OpenAIDevs Input tokens for o1 are calculated the same way they're calculated for 4o. Both models use the same tokenizer:
1
0
14
Full house at E-Commerce Day in Hong Kong organized by Stripe, Easyship, Shopify, and Google #ecommdayhk
1
3
13
This is excellent advice on building AI applications iteratively. @colintjarvis and team do some of the most impressive work at OpenAI and have been a key part of making our largest users successful. Highly recommend!.
Optimizing LLMs for accuracy is hard - despite opening our talk at DevDay in November last year with those same words, achieving the right level of accuracy for production remains a challenge. I’ve tried to collect my learnings from the last year into
1
0
12
@atroyn @OpenAIDevs Not at the moment. This is a neat idea and we'll explore it for future models!.
1
0
11
We're seeing many examples of users combining our reasoning models with 4o/4o-mini to unlock new capabilities. Here's an excellent example by @GrabSG showing how you can use o1 to do better data labeling for your SQL generation pipeline.
Cool way to use OpenAI o1 we haven't seen yet: Lead data scientist Christian Coffrant was part of a team at @GrabSG that combined GPT-4o (for writing SQL) and o1 (for data labeling) to build an AI data analyst called Jarvis. It looks through a vector database, retrieves examples,
2
0
12
@btibor91 @OpenAIDevs This is very helpful feedback. Feel free to DM me with any examples you feel comfortable sharing (chatgpt share links / api request ids).
1
0
11
++ this is my top personal use-case for o1. it’s really good at completing or rewriting emails or docs in a way that preserves my tone and writing style.
new learning from this week:. o1 is extremely good at copy editing. no other LLM (including Claude!) is any good at this. gigantic performance improvement.
1
0
11
@itsurboyevan @OpenAIDevs We found in our early testing that putting too many less-relevant chunks into the context could result in worse performance and high reasoning token usage. We may be able to improve this in future versions, but for now, we recommend using the fewer & more relevant chunks.
2
0
10
@swyx Great question! It does get back on track after it sees the token matching starting to converge between the prediction and the model output. The current threshold to get back on track is a 32 token match. We have some future optimizations planned that will make this even lower.
3
0
9
Overheard at Somerset MRT: we’re blockchain-ing the cloud to make AWS instances tamper free. We sign the instances and put them on a blockchain. We’re also doing an ICO. #thingsthatcanbesolvedwithoutblockchains.
0
1
7
So excited about the results here -- 4o-mini is the cheapest model in it's category and is outperforming models several times it's size. Best part of all, you can finetune 4o-mini starting today, for *free* (up to 2M training tokens/day):
Exciting Chatbot Arena Update -- GPT-4o mini's result is out!. With 4K+ user votes, GPT-4o mini climbs to the top of the leaderboard, now joint #1 with GPT-4o while being 20x cheaper! Significantly better than its early version ("upcoming-gpt-mini") in Arena across the boards.
1
0
9
@tillstr__g_l @OpenAIDevs It's hard to predict exactly where o1 and 4o prices will land in the future, but we're confident per-token prices for some o1 models will go lower. We’re committed to continuing the trajectory of offering more intelligence per dollar!.
0
0
9
We've added polling helpers to the Node + Python SDKs for Run objects in the Assistants API. If you're not streaming, you can now use:.- Node: threads.runs.createAndPoll.- Python: threads.runs.create_and_poll. (and optionally modify the poll interval to poll faster/slower).
0
1
9
Fortunately, there's a multi-billion $ industry of infra companies popping up to make dapps faster, simpler, more secure. The AWS / RoR of the decentarlized world are yet to be built. Some of my favorites infra players so far include: @0xPolygon, @ChainLink, @magic_labs!. ~fin~.
0
1
8
The Metaverse Primer is 💯. I've been naively dismissive about the Metaverse because of how lackluster my experiences with Oculus / AR / Decentraland have been so far. But there's something about the way Matthew clearly lays out the challenges that makes me optimistic about it.
0
1
8
Welcoming @athyuttamre back on twitter with this banger.
We're starting to roll out improved error messages in the OpenAI API, including typo detection, reliable error codes, and actionable suggestions:
2
0
8
@Kwebbelkop @OpenAIDevs we're early in our journey here but we're excited to figure out how to bring fine-tuning to these models. definitely on our roadmap!.
0
0
8
One of the last things I worked on at @stripe -- go @jiegantic and team! 😍.
We’re adding support for @UnionPay_Intl card payments for businesses in over 30 countries. Along with WeChat Pay and Alipay, Stripe businesses can now sell to billions of Chinese customers around the world. 🌏
1
0
7
Can’t believe I’m saying this but I miss lurking on Stripe email lists so much. @TechEmails has lately been providing some much needed relief!.
0
0
6
Just 2 weeks into exploring #DeFi & already amazed at the number of new concepts I've learnt (despite having spent >5 years in "CeFi"). Favorites so far:. - 📈Quadratic funding @gitcoin. - 💰Liquidity pools & AMMs @Uniswap. - 🌩️Fast, free, decentralized settlement @lightning.
1
0
6
Literally everyone on the team is working to ensure we continue to deliver for our customers. Lucky to have customers like @BeMyEyes.
So. About @OpenAI. I've been bombarded by sales calls from rival LLM companies seeking some opportunistic business wins. That's fair game and business is a contact sport so I get it. And we were already evaluating backup providers as a hedge before the drama. But the reason we're.
0
0
7
I think I subconsciously have *more* trust in startups with female founders (not sure why). Wanted to signup for @birchbox and having had bad experiences with subs before, decided to do 2 mins of due diligence on Google — felt relieved when I saw female founders ¯\_(ツ)_/¯.
1
0
7
It honestly feels like there are fewer homeless per capita in Bangalore than in SF. It’s actually hard to find a homeless person on the streets here.
2
0
5
@rinte0321 @OpenAIDevs We're excited about this and it's on the roadmap! No timelines to share yet unfortunately.
1
0
7
@billyuchenlin @OpenAI Exciting results! I work on the OpenAI API. Do you have the request IDs of the empty string responses? We would love to take a look and figure out what’s happening here.
1
0
7
@APompliano This list would be completely different if you removed car ownership from the benchmark. Compared to LA/SF, way fewer in SG/CH/NYC/HK want / need cars because of excellent public transportation infra.
0
0
2
Seeing Claire grow Stripe and being a part of her org was *the* highlight of my time at Stripe -- she's incredible, and I'm super excited about this!.
Scaling People: Tactics for Management and Company Building by @chughesjohnson is a practical guide to building a company’s operating systems and scaling the most important resource a company has: its people. Available for preorder now: 📈
0
0
6
nothing better than spending your saturday giving the canadian government 10 years of travel history with dates (and entry + exit stamps for proof!) so that you can help boost their tourism sector.
4
0
6
In love with the new page! Kudos to whoever worked on this around getting all of this information together in such a neat manner.
0
0
5
@BenPielstick @OpenAIDevs Hard to give an exact timeline but we are working hard to increase rate limits. We'll keep you all posted!.
0
1
6
The situation on the ground in India is really bad — several friends have had their relatives in India die during this second wave. To make things worse, funeral homes are overbooked and they can't travel back to say goodbye.
1
2
1
@eaftandilian @OpenAIDevs thank you for all the early feedback and testing! can't wait to see this live in Github 🙌.
0
0
6
The amount of crypto spam on Twitter (message requests + mentions) is ridiculous. It’s also really cumbersome to report spam. Needlessly long multi-step process when you select ‘Report Tweet’.
6
0
5
@SundaysInHell @patrickc According to @dps's sundial generator, it looks like sundails are quite a mess in Singapore:
0
0
4
Really enjoyed this overview of how Constellation Software operates. Great tidbits around keeping teams small, autonomous, & focused on customers. It's surprisingly hard to find content on $CSU but @Austen has compiled all their shareholder letters here:
I wrote an Operating Manual for Mark Leonard, CEO of Constellation Software. Mark has bought 500+ niche software companies over the last 25 years. And turned $25 million into a $32 billion empire. This is how he makes it work.
1
0
5
Last week was all about learning what makes scaling @ethereum hard and what folks are doing about it. Over the next couple of years, the scaling strategy involves moving transactions 'off-chain' (i.e. from Layer 1 ➡️ 2).
2
0
5
@fyfrimpong Beam is a remittance service: people in EU send us BTC & we instantly convert to mobile money in GH. Kitiwa sells BTC in GH.
0
2
3
@apoorv03 @OpenAIDevs oh hi :) yes you can. we'll add batch api support once the rate limits are higher.
0
1
5
Our models are simultaneously getting much more capable 📈 and much cheaper 📉.
Big developer updates today.- new embeddings (up to 5x cheaper, more capable, truncatable).- update GPT-4 Turbo model. Plus another major price cut for GPT-3.5 (3rd one this year!). 50% cheaper for prompts, 25% cheaper for outputs.
1
0
5
@Hemkunt_Fdn Thank you for everything you’re doing. I’m trying to make a donation from overseas and am using a remittance company (DBS remit). To complete the transfer, they’re asking what type of bank account you have: NRO, NRE, Savings, Current, or Cash Credit. Please lmk!.
3
0
4
tfw when you meticulously plan your road trip on @googlemaps but end up closing the app by mistake and losing the whole thing in the middle of your nowhere with no data connection.
1
0
3
at least canada gives indians 10-year multi-entry visas. applied for a swiss visa recently for a 7-day trip and they generously gave me a *9-day* single-entry visa. feel like extending your trip? 🙅♂️🙅♂️🙅♂️ you're brown!.
0
0
3
Sold mobile phones at Sim Lim Square! $6 / hour.
What was your very first job and how much were you paid an hour?.
1
0
3
@ReporterWeather @OpenAIDevs This would make for a neat feature! Not possible with the current models, but we'll look into supporting this for future ones.
1
0
4
Also! o1 rate limits are as high as 4o now + available to T3. Happy reasoning 🚀.
🪴 OpenAI o1 API—we’re expanding access to developers on usage tier 3 and increasing rate limits (to the same limits as GPT-4o) so your apps can be production-ready. Tier 5.o1-preview: 10,000 requests per minute.o1-mini: 30,000 requests per minute. Tier 4.o1-preview and.
0
0
4
I really hope onename.io becomes mainstream and more #bitcoin wallets adopt it (like rushwallet!) http://t.co/RJ7mZZFY8l.
0
0
4
.@StripeSingapore is hosting our very first Ruby SG meetup tonight! Sign up and come on by tonight to check out our ✨ new office
0
0
3
GPT-3 life hack I've been using for the past couple of days:. ```.${email body}.------.take a pass at the email above and make it clearer. keep it informal and friendly. ```. Paste result and send.
1
0
3
Finally! So much thought & rigor by the APAC team here — particularly around bringing the best FPX integration to market 💸.
Stripe is now available in Malaysia. 🇲🇾
0
0
2
What did you get done this week?.
What an insane week for Elon:. - Tesla AI Day.- Q3 Production & Deliveries .- Twitter vs Elon.- S&P investment grade upgrade.- SpaceX human launch to station.- Starlink launch .- Tesla Semi production . I don’t know how this guy does it. Make sure you stay healthy @elonmusk!.
2
0
2
@Amegaxi buying something from eBay with the help of @KitiwaBTC and @joinsnapCard. He just bypassed Paypal in Ghana! http://t.co/v7vls5Ll8X.
2
6
3
0
0
2
Stripe has partnered with a small cohort of local VCs and accelerators on a new research project into the “startup s…
0
0
3
I sometimes look at my ability to purchase things with my credit card as a type of insurance - i.e. if the product or service is not / poorly delivered, I always have the last resort option of filing a dispute.
1
0
2
Proud to have had the opportunity to work with the kickass team at Govtech! Some of the best product / eng minds in SG. 💯
‘Biggest mistake’ for Not putting prototype in users’ hands faster, says @GovTechSG's Li Hongyi #smartnation
0
0
2
ChatGPT is cool and all but when are we going to have AI that's able to visually align logos on a slide.
1
0
3
If you're looking for a simple + effective + impactful way to help out, consider donating to @india_oxygen here: . This effort is bootstrapped by a group of founders in the Delhi area and they're transparently sharing updates on their Insta & Twitter!.
0
0
2
@dus0l @btibor91 @screenstudio @pie6k yes! used it for the first time for this. love what you've built @pie6k.
0
0
2