shyamal
@shyamalanadkat
Followers
8,700
Following
977
Media
59
Statuses
2,579
applied ai @openai . 🇮🇳 @dukeu
San Francisco Bay Area
Joined April 2019
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
台風10号
• 224132 Tweets
Oasis
• 195132 Tweets
Michael
• 175159 Tweets
台風の進路
• 128690 Tweets
Racing
• 106597 Tweets
श्री कृष्ण
• 78468 Tweets
Gökhan Akkan
• 56832 Tweets
Independiente
• 43745 Tweets
Noel
• 38338 Tweets
Fortaleza
• 32834 Tweets
Allan
• 30405 Tweets
#يلو_ترعي_o15
• 28421 Tweets
FLOR ESTAMOS CON VOS
• 25896 Tweets
Pedro Raul
• 25126 Tweets
#KamalaWinsWithJoy
• 20614 Tweets
Erick
• 20192 Tweets
Millonarios
• 19195 Tweets
Wesley
• 17755 Tweets
Angélica
• 17343 Tweets
Bragantino
• 14966 Tweets
Seabra
• 13650 Tweets
Xuxa
• 13085 Tweets
Ramiro
• 12900 Tweets
Luiz Araújo
• 12681 Tweets
Yuri Alberto
• 12541 Tweets
#Domingão
• 11549 Tweets
कन्हैया लाल
• 11504 Tweets
🩶
6
20
790
on my recent visit to india, i met a super sharp 19-year-old student from a remote town in gujarat, india. surprisingly, he traveled for 11 hrs each way to meet me for a 3 hr long discussion about agi, consciousness, creativity, anomalies in our approach to RL, spatial computing,
23
32
620
exploring k-means clustering on the Iris dataset using code interpreter chatGPT plugin ✨
22
94
559
we also dropped the chatgpt retrieval plugin that enables semantic search and retrieval of personal or organizational documents. s/o to
@IsaFDev
.
11
57
414
here are some basic → advanced strategies that I see for optimizing RAG implementations these days:
basic
1. using effective prompt eng, templating, and conditioning. eg: “given the context information and no prior knowledge, answer the query..” etc. ok, we've all done some
7
45
356
today, we’re releasing the Assistants API, our first step towards helping developers build agent-like experiences within their own applications.
inspire. build. dream.
6
25
322
The LLM landscape is increasingly starting to look something like this: (1) Model layer (e.g gpt3, cohere)(2) API bindings for access (e.g openai-python) (3) Infra layer for prompt chaining/model switching (e.g.,
@LangChainAI
,
@dust4ai
) (4) Next gen AI-first apps. what's missing?
13
19
270
shipping AI without calling it AI is underappreciated. it brings back focus on users and value; no hype, just utility that users gradually rely on. often the best solutions hide complexity behind simple interfaces.
20
33
270
updated guide for Q&A with GPT and embeddings-based search OR "how do I integrate with ChatGPT using my own custom data" - ; s/o to the amazing
@sandersted
5
26
255
hot take: RAG has become unnecessarily too complex and way too commercialized/productized; we should really work on simplifying it and build cleaner, better abstractions.
31
11
255
new guide for building a tool-using agent with LangChain OR "how do I augment ChatGPT with my own custom data and external tools/APIs" -
; s/o to
@colintjarvis
3
54
245
some of the most talented human agents hacking on AI agents at the
@scale_AI
genAI hackathon. i love san francisco!
11
13
217
recently got a call from a cousin who said the GujRERA (gujarat real estate regulatory authority) portal was taking about 20-30 seconds to load when he was trying to get a simple question answered. so, i quickly prototyped an openai +
@embedchain
powered assistant that provides
12
20
204
LLMOps is young but growing fast; key areas include:
(1) prompt management & evals (chaining, audit, a/b testing, etc) (2) simplifying finetuning & embeddings (3) agent integration/action execution (interfacing w/ external world) (4) analytics & observability - cost, latency, etc
4
12
187
few verticals & companies working with LLMs i’m excited about:
1/ legal tech startup
@LexionAi
is using GPT-3 to help lawyers write summaries and suggest edits.
5
23
161
Key moats for LLM/AI-first applications, in order of importance: (1) Proprietary data and fine-tuning (2) Great UX + one that instills a sense of trust and reliability (3) Distribution and GTM (4) Network effects and community (5) Breadth and depth of integrations
What else?
10
20
150
at the chatgpt API hackathon today in SF organized by
@rachel_l_woods
and
@nonmayorpete
in ~3 days. the energy is off the charts with over 200+ in attendance. can’t wait to see the demos 🔥
1
6
132
i don't say this enough but i love openai. it's absolutely been the highlight of a lifetime to work here, drive the kind of positive impact, help shepherd the technology to help humanity and progress towards agi, and collaborate with some of the most talented, hardworking,
6
7
123
OPENAI OPENAI OPENAI OPENAI
DEVDAY DEVDAY DEVDAY
DEVDAY OPENAI OPENAI OPENAI
DEVDAY DEVDAY DEVDAY
OPENAI OPENAI OPENAI
OPENAI DEVDAY DEVDAY
5
9
115
silicon valley at its best at
@scale_AI
’s generative AI hackathon this saturday in SF.
s/o to
@russelljkaplan
and his team for making this happen.
5
15
113
Check out OpenAI's new Production Best Practices guide! It covers a range of topics to help developers transition from prototype to production. Let us know what you think!
3
21
112
"prompting LLMs is almost like subtractive sculpting; each additional word in the prompt eliminates some possible worlds and defines the task more clearly" -
@goodside
3
11
108
1/ practical tactics to reduce LLMs from hallucinating🧵: during training, LLMs auto-regressively predict the next token in a sequence. despite their impressive performance, LLMs sometimes generate outputs that are factually incorrect/ nonsensical, known as hallucinations.
4
14
104
At OpenAI DevDay,
@sherwinwu
and I had the pleasure of giving a talk on 'The New Stack and Ops for AI', where we discussed a new framework to navigate the unique considerations for scaling non-deterministic apps from prototype to production. if you couldn't join us live, you can
5
9
95
next couple of years will present a steep tech curve - AI-first companies disrupting entire industries with beautiful language interfaces, LLM apps, and novel interaction paradigms resulting in 10x productivity/efficiency gains & driving the marginal cost of knowledge work down.
6
15
85
sharing some insights that can help founders and builders navigate the landscape, enabling AI-enabled and AI-native products/experiences that can amplify humanity:
1
11
78
Thin layers around LLM APIs aren't enough to gain a competitive edge in AI-first apps. To win, you need deep integrations and optimized workflows that solve real problems with the scalability and efficiency that wasn't possible pre-GPT.
3
13
79
excited about the potential for LLMOps & LLM tooling (things like prompt testing & experimentation at scale, vector db & ops, data to embeddings, no code, etc). with foundational models in place, tooling & integrations will make them more reliable & easier to operationalize.
6
8
77
new guide on OpenAI API rate limits and how developers can batch request / handle retries: , s/o to
@OfficialLoganK
1
14
77
gpt4-o api
• 2x faster
• 50% cheaper
• 5x higher rate limits
3
6
78
"Getting fired from Apple was the best thing that could have ever happened to me. The heaviness of being successful was replaced by the lightness of being a beginner again, less sure about everything. It freed me to enter one of the most creative periods of my life."
- Steve
3
5
75
looking to connect with a few IITians to discuss the current state of education in india. if you're an IIT student / alum, would love to hear your insights. feel free to DM or reply below.
26
9
76
0/ compiling some openai function calling API tips/pointers from few developers:
1
16
70
humbled by the shipping intensity, consistency, and velocity at
@openai
. it continues to amaze me how the team just gets it done.
4
2
71
Creating a ChatGPT plugin in 3 steps:
1/ build an API
2/ document the API in the OpenAPI yaml or JSON format
3/ create a JSON manifest file that will define relevant metadata for the plugin
example todo list plugin:
2
13
67
i love openai, and everything we stand for, the people, the tech, the problems we're solving, the impact we're having, and staying committed to learning & growing along the way.
1
3
64
Memes, Brainstorm, Knowledge, Identity
2
6
61
focus on doing a few things exceptionally well and be committed to those.
0
5
60
👑
2
2
58
obsessed with AI startups accelerating knowledge workers rn.
7
3
60
@paulg
the dealer's profits rely on the liquidity and turnover of the artwork, not on its potential long-term appreciation; it’s not necessarily about the highest future value, but the optimization of transaction frequency and profitability
2
2
58
you really don't know/understand what hypergrowth feels like unless you're part of it day in, day out. the intensity and pace are hard to grasp from the outside.
3
3
57
i love cursor. absolutely great team and delightful product.
Cursor users can now use the new gpt-4-turbo model. We’ve observed improved reasoning on complex tasks.
Example comparison between gpt-4-1106 and the new gpt-4-turbo below:
23
53
547
2
4
57
prompt engineering is such a fascinating & creative field that combines NLU, domain knowledge, and human intuition to elicit the best possible responses from LLMs. it will evolve with LLMs, requiring new methods and tools to harness their full potential.
3
7
54
moved from chrome to arc.
swapped vscode for cursor.
shifting from docs to notion.
thanks for shipping, innovating, and reshaping my stack
@browsercompany
@anysphere
and
@NotionHQ
7
3
53
AI-first apps will transform how we work & collaborate over the next 5 years, making knowledge work and intelligence more accessible & affordable. Note-taking & copyrighting are just the beginnings. New interfaces, CRMs, tax prep copilot, research assistants are all fair game.
3
8
53
introducing gpt-4o-mini!
15 cents per 1M input tokens and 60 cents per 1M output tokens, scores 82% on MMLU and currently outperforms gpt4 on chat preferences
announcing our new small model gpt-4o-mini! smarter than Claude Haiku and Gemini Flash, and 30x cheaper than gpt-4o
20
41
533
4
1
53
700+ and counting
About 650 / 770 signed at this moment. As people start waking up, more will come. All the efforts started after 1:30 AM, 500+ within two hours and all of this after 2 crazy days with very little sleep.
156
594
5K
0
2
52
re: finetuning LLMs
0/ you don't always need it. try everything else first.
1/ lot of data + limited compute -> go for PEFT
2/ limited data + lot of compute -> full fine-tuning
3/ limited data + limited compute -> `import openai` ;)
2
4
48
the ‘low key research preview’ over the past year evolved from a toy to a tool, and now to a capability that millions depend on.
happy birthday,
@ChatGPTapp
; we are just getting started.
a year ago tonight we were probably just sitting around the office putting the finishing touches on chatgpt before the next morning’s launch.
what a year it’s been…
2K
3K
44K
0
2
47
logprobs are back! you can use it as a likelihood function or get confidence scores on classification tasks, enabling more interpretability in outputs.
By popular demand, log probabilities are now in the Chat Completions API!
With logprobs for top output tokens, you can build features such as better autocomplete, classification, keyword suggestion, or assess the model’s confidence in its predictions.
65
198
2K
4
1
47
the most important thing I have realized this year: life is truly and should be centered around people you love and people you serve. everything else is nice to have.
3
3
46
structured outputs in the API! also the updated gpt-4o model is 50% cheaper for input, 33% cheaper for output, and is available immediately.
3
5
44
recursive prompt chaining is a powerful tool; leveraging LLMs to evaluate your completion outputs and further assessing those evaluations lets you fine tune results for maximum efficacy and robustness across many use-cases.
2
3
44
for founders & investors: join
@openai
startup community and stay connected with our team and learn about upcoming events and technical resources. we have lots of great things cooking:
0
5
45
hello, sora
Introducing Sora, our text-to-video model.
Sora can create videos of up to 60 seconds featuring highly detailed scenes, complex camera motion, and multiple characters with vibrant emotions.
Prompt: “Beautiful, snowy
10K
32K
138K
3
1
42
cheap, abundant, and customizable intelligence at scale has to be this year's theme.
1
5
44
the
@full_stack_dl
LLM Bootcamp may be the best event taking place this weekend in the bay area. kudos to everyone involved in making it happen!
0
4
43
huge thank you to the entire openai leadership team for making this happen🩵 one team, one mission.
0
0
43
the trojan horse for AI is not anthropomorphism but intelligence augmentation. rather than building bots/integrations that aim to pass as human, the best founders will create tools that feel like natural extensions of our own abilities. the real goal is leveraging computation to
3
3
40
some of the most valuable AI first companies will create entirely new UX paradigms and interaction models powered by elegant natural language interfaces
3
6
40
can't wait to see a world where we don't have to worry about some llm api rate limits, incremental costs, and inference latencies, where we continue to drive down the cost of compute + intelligence. abundance will unlock tremendous value and adoption, even with current
0
2
39
@sama
life is truly and should be centered around people you love and people you serve. everything else is nice to have.
1
3
37
agreed. shipping stuff at openai is so inspiring, exciting and brings the best out of silicon valley
I don’t think people comprehend how much work goes into OpenAI launches, it’s exhausting and yet so inspiring.
48
14
497
2
0
37
dev ex is everything.
OpenAI's moat is it not taking 17 clicks to find your API key
15
14
306
3
2
37
outlining some practical measures for limiting LLM hallucinations in this post -
1
11
38
as the doubts about creating a moat with LLMs fade, we recognize that the platform does not guarantee equality. it's like gaming consoles being around for decades, yet game developers are still competing in creating the most immersive games. this highlights the importance of
2
8
35
exciting updates
- function calling capability in the api with updated & more steerable versions of gpt-4 and gpt-3.5-turbo
- 16k context version of gpt-3.5-turbo
- 75% cost reduction on embeddings api
- 25% cost reduction on input tokens for gpt-3.5-turbo
3
4
34
gpt4 can act as an effective discriminator/verifier to gpt 3.5 outputs. for eg - by asking 3.5 to answer a question and then asking gpt4 to verify the validity/truthfulness of the answer given the context, we can achieve a higher level of confidence.
2
4
33
The three stages of LLM developments:
1.0: Capable of generating original text and reasoning about it.
2.0: Able to evolve, refine its output, and acquire new abilities to act rationally.
3.0+: Can design its own actions/capabilities to interact with the external world
4
1
35
your users don’t have to know you’re building with LLMs just like they don’t have to know you’re building with javascript; intelligence is value. and necessary, not just an accessory.
2
3
33
we’ve come a long way. code interpreter available to all chatGPT plus users over the next week - a preview into the future of work.
exploring k-means clustering on the Iris dataset using code interpreter chatGPT plugin ✨
22
94
559
1
4
33
even in a world where developers have cheap access to very capable AI systems, i think the most challenging part of this game will still be building the integrations, infrastructure, data integrity, and essentially the "glue" between intelligence/software 2.0 <> 1.0. developing
5
2
32
chatgpt plus and gpt4 are now available to everyone in India.
this is a huge milestone for us and it’s going to be really exciting to see the possibilities 🇮🇳 ❤️
1
4
32
few areas within LLM space I’m personally excited about:
(1) horizontals - tooling/infra, no/low code, finetuning, prompt chaining + retrieval, actions, experimentation frameworks
(2) verticals - copilot for x, AI-native CRMs, edtech, health, enforcement/safety. Lots to explore!
0
3
30