rahul @rahulgs Twitter profile | Pikagi

Pikagi

rahul

@rahulgs

2,993

Followers

968

Following

67

Media

325

Statuses

leading ai @tryramp . cofounder cohere (acq ramp)

NYC

https://t.co/h4WckWnm0k

Joined December 2014

Don't wanna be here? Send us removal request.

Pinned Tweet

@rahulgs

rahul

1 year

Problem: Getting LLMs to output valid JSON in the format you want is hard Solution: ONLY generate values, feed model with keys and JSON structure. Constrain outputs with custom sampling New project: Jsonformer: A Bulletproof Way to Generate Structured JSON from Language Models!

Tweet media one

43

135

1K

Last Seen Profiles

@W650_mihoko

@Amjaklo

@xcou80

@ReifsteckC72905

@STimohty49617

@JShakema54769

@RileyMassey1511

@mFitzroy55

@MalkaNicki60792

@Soma_Safe_

@cukienaknikmati

@charlena52727

@NmA20160

@KaleJustin83953

@DarriusTae4809

@GaytonK48371

@CoachMooreSTL

@StonecoldSxnick

@2la90

@HYURNJlN

@SXMGOV

@Alphatled08

@felipejungm

@onosatori

@Moh_Mirghany

@BinorRaja

@KJenniferm65276

@yeahytk

@Coach_Russo

@KAppolonia32370

@Kana_terao

@TheLanceTaylor

@SPINALL

@NAPALM_DEATH

@imlama_1

@Aseia77

@rahulgs

rahul

4 months

at @tryramp we use LLMs to find the 5 most valuable mins of audio from the 1000+ customer calls we make every day narrated by TTS + compiled into a 5 min podcast sent to the entire team

@stevekrouse

Steve Krouse

4 months

as a product owner it'd be nice to have an llm summary of everything my users did yesterday calling out cool success stories or troublesome error states i should reach out to debug has anyone tried such a thing? i am thinking about prototyping it with public val town data

43

10

320

22

31

551

@rahulgs

rahul

4 months

🤔 extracted the full ~5000 token claude3.5sonnet system prompt: this is a great template for function calling / tool use notes: artifacts: seem to be a fully in-context abstraction, model not finetuned for it allowed types:

Tweet media one

15

58

423

@rahulgs

rahul

3 months

introducing 🪄genweb: the first software 2.0 web framework 🪄 genweb is a new way of building web apps: instead of a frontend and backend codebase, an LLM is the backend and the frontend it interprets user actions and dynamically generates UI in real-time welcome to the

@karpathy

Andrej Karpathy

4 months

100% Fully Software 2.0 computer. Just a single neural net and no classical software at all. Device inputs (audio video, touch etc) directly feed into a neural net, the outputs of it directly display as audio/video on speaker/screen, that’s it.

596

739

8K

20

25

362

@rahulgs

rahul

2 months

we're hiring full stack engineers to work on llms at ramp if interested, dm me with examples of real things you've built come work on real deployments and learn how to drive enterprise value we're a small and mighty team what we've worked on in the last year: - multi-step

19

21

338

@rahulgs

rahul

2 years

🎉 new project: Clarity! A reading app that offers a fresh approach to consuming text. Instead of the traditional linear reading style, Clarity allows you to read depth-first, diving into the details that interest you most.

16

27

312

@rahulgs

rahul

4 months

generates CoT tokens within <antThinking> tags, hidden from user on the server

Tweet media one

17

11

239

@rahulgs

rahul

7 months

got Devin to fix bugs in OpenDevin

Tweet media one

8

9

219

@rahulgs

rahul

1 year

Excited to announce that we're joining forces with one of our customers, @tryramp , where we will help build the future of AI + finance

@bayareawriter

Mary Ann Azevedo - out of office

1 year

Exclusive: @tryramp makes its 2nd acquisition, scooping up , which has built out an AI-powered customer support tool.

2

17

144

12

3

152

@rahulgs

rahul

11 months

I kept having to debug prompt issues with open models So I built OpenAI's Tokenizer page for all tokenizers on HuggingFace: Llama, Mistral, GPT2, MPT, Persimmon, T5 etc check it out here:

Tweet media one

8

22

147

@rahulgs

rahul

3 months

jsonformer + openai! "deterministic, engineering-based approach to constrain the model’s outputs to achieve 100% reliability"

Tweet media one

@rahulgs

rahul

1 year

Problem: Getting LLMs to output valid JSON in the format you want is hard Solution: ONLY generate values, feed model with keys and JSON structure. Constrain outputs with custom sampling New project: Jsonformer: A Bulletproof Way to Generate Structured JSON from Language Models!

Tweet media one

43

135

1K

3

8

145

@rahulgs

rahul

3 years

me: can you pass the water homebrew: updating homebrew

2

1

72

@rahulgs

rahul

1 year

My favorite thing to do on Modal - running massively parallel GPU finetune jobs At Ramp, we’ve trained hundreds of LLMs *at the same time* without the infra hassle - Modal allows us to move insanely fast (1/2)

@modal_labs

Modal

1 year

Modal is generally available today, and we also raised a Series A!

38

82

690

3

6

128

@rahulgs

rahul

1 year

that’s a lot of stars 🤯

Tweet media one

4

1

116

@rahulgs

rahul

2 years

look at this linkedin dm i just got

Tweet media one

3

4

103

@rahulgs

rahul

3 months

Shoutout to Ramp engineer Andrew Gu who was on the coaching staff. Congratulations to Team USA for winning IMO 2024!

@johncoogan

John Coogan

3 months

Congratulations, welcome to the Ramp engineering team.

Tweet media one

7

9

464

0

4

90

@rahulgs

rahul

11 months

the netflix documentary is gonna go crazy

Tweet media one

1

2

86

@rahulgs

rahul

4 years

so pumped to share the news, and ...we are just getting started! how we got here ↓ (thanks @sarahintampa for the awesome coverage!)

Tweet card media

Cohere raises $3.1 million for its remote control solution for web apps | TechCrunch

Existing remote desktop solutions like LogMeIn and TeamViewer can be complicated to set up and use, and can feel dated. A new startup called Cohere, now

5

6

80

@rahulgs

rahul

2 months

"jobs not finished" - @eglyman

Tweet media one

2

1

81

@rahulgs

rahul

1 year

Generate perfect schema-conforming JSON, every time:

Tweet media one

3

7

80

@rahulgs

rahul

10 months

Tweet media one

2

0

77

@rahulgs

rahul

1 month

with the o1 release, reminder that has been using thinking tokens for several months now

Tweet card media

Introducing OpenAI o1

Introducing OpenAI o1

@rahulgs

rahul

4 months

generates CoT tokens within <antThinking> tags, hidden from user on the server

Tweet media one

17

11

239

1

2

69

@rahulgs

rahul

1 year

Jsonformer supports a subset of JSON Schema, including number, boolean, string, array, and object types. It's built on top of the HuggingFace transformers library, making it compatible with any model that supports the HuggingFace interface. Try it —

Tweet card media

GitHub - 1rgs/jsonformer: A Bulletproof Way to Generate Structured JSON from Language Models

A Bulletproof Way to Generate Structured JSON from Language Models - 1rgs/jsonformer

4

6

66

@rahulgs

rahul

3 years

Honored to be part of the 2022 Forbes 30U30 list with my cofounder @yunyu_l for @CohereHQ

Tweet media one

4

2

62

@rahulgs

rahul

1 year

New complex schema generation example live With just a tiny 3b model (databricks/dolly-v2-3b)

Tweet media one

Tweet media two

2

5

57

@rahulgs

rahul

6 months

i achieved 100% accuracy on 0.007% of swe bench

@kevinlu1248

Kevin Lu

6 months

Sweep achieves 15.7% on SWE-bench! Hi everyone, we’re building Sweep, an open-source AI developer that handles the easiest 30% of software tasks. We’re thrilled to announce our results on SWE-Bench! We evaluated Sweep on a random 10% subset of the data. Sweep correctly

Tweet media one

21

23

267

3

0

50

@rahulgs

rahul

1 year

Generating JSON is probably a common enough use case that hosted model providers should probably support an JSON only API thoughts? @gdb @aidangomezzz @AnthropicAI

3

2

49

@rahulgs

rahul

1 year

I finetuned an LLM on all my iMessages, try it on yours! releasing code with sql queries, data processing, finetuning with PEFT and a chat CLI

Tweet media one

3

1

48

@rahulgs

rahul

17 days

what did you get done in the last hour

14

0

47

@rahulgs

rahul

2 months

jobs not finished

@tryramp

Ramp

2 months

we have work to do

Tweet media one

7

1

88

5

0

44

@rahulgs

rahul

1 year

huge

@LangChainAI

LangChain

1 year

❓How to get models to generate structured output? JSONFormer (by @rahulgs ) and RELLM (by @mattrickard ) are two novel approaches for this, now with (experimental) integrations to LangChain JSONFormer Integration RELLM Integration

Tweet media one

Tweet media two

11

42

295

0

5

41

@rahulgs

rahul

3 months

openai: with structured mode vs without in my benchmark, structured extraction mode is 13% slower, samples about the same number of tokens code:

Tweet media one

6

5

42

@rahulgs

rahul

2 months

💪

Tweet media one

4

2

41

@rahulgs

rahul

1 year

Problem: Generating structured JSON from language models is challenging. Current approaches like prompt engineering, fine-tuning, and post-processing often fail to produce syntactically correct JSON.

Tweet media one

2

2

39

@rahulgs

rahul

3 years

when i was at superhuman it would bother me immensely when people called us SuperHuman we've come full circle

Tweet media one

4

0

39

@rahulgs

rahul

7 days

Tweet media one

@isamlambert

Sam Lambert

7 days

Uber runs 16,000 MySQL nodes. Actual scale.

5

55

406

4

0

140

@rahulgs

rahul

7 months

it’s time to cook

1

3

35

@rahulgs

rahul

6 months

Tweet media one

0

1

36

@rahulgs

rahul

3 years

Worked long and hard on this one - incredibly hard to get this just right!

3

0

35

@rahulgs

rahul

3 years

what the

Tweet media one

2

1

35

@rahulgs

rahul

1 year

Solution: Jsonformer: A wrapper around HuggingFace models that only generates content tokens and fills in fixed tokens during the process. This makes it more efficient and bulletproof than existing methods

2

1

34

@rahulgs

rahul

3 years

A++ customer support from @will_ye_ sign up @CohereHQ and we'll write you a haiku

Tweet media one

3

1

33

@rahulgs

rahul

2 years

Source: Based off of @andy_matuschak 's amazing Evergreen notes and @OpenAI 's "Recursively Summarizing Books with Human Feedback" () And huge thanks to @yunyu_l @thesephist for feedback

2

0

32

@rahulgs

rahul

3 years

literally everyone under the age of 25 who has invested in Cohere has asked if they can Venmo me the money

4

0

33

@rahulgs

rahul

2 months

when I was making this graphic in Figma @yunyu_l told me to change to the latex font so it looks more academic

@Teknium1

Teknium (e/λ)

2 months

Thought this said jensenformer

Tweet media one

15

7

159

3

0

31

@rahulgs

rahul

9 days

@shaig

0

0

30

@rahulgs

rahul

7 months

this is just the beginning, excited to be a supporter

@cognition_labs

Cognition

@cognition_labs

7 months

Today we're excited to introduce Devin, the first AI software engineer. Devin is the new state-of-the-art on the SWE-Bench coding benchmark, has successfully passed practical engineering interviews from leading AI companies, and has even completed real jobs on Upwork. Devin is

5K

11K

45K

0

0

29

@rahulgs

rahul

3 years

me: looks like a 30 minute feature, quick n easy also me 5 hours later:

Tweet media one

1

0

29

@rahulgs

rahul

3 years

🤯 our most requested feature is out!

@CohereHQ

Cohere

3 years

Announcing Cohere Voice! The same frictionless experience, now with audio and video. It’s that easy. 🎤📹

2

5

62

0

0

29

@rahulgs

rahul

4 months

here’s an example (voices and quotes altered)

3

1

29

@rahulgs

rahul

2 years

didn't get access to copilot x yet so I wrote my own with gpt4 try bropilot, a rust cli that helps you write terminal commands

Tweet media one

0

2

27

@rahulgs

rahul

4 months

this is how you talk to your users at scale

0

0

27

@rahulgs

rahul

11 days

“In 15 words: deep learning worked, got predictably better with scale, and we dedicated increasing resources to it.” - Gandhi

@calixo888

calix huang

11 days

one of the hardest of parts of building a good agentic UX is integrating to the user's context. automation only works when we can make intelligent decisions without requiring a user to put in extra work. proud to have led this project alongside many others at @tryramp 🤝

6

3

71

2

0

27

@rahulgs

rahul

1 month

agree, 100% a mistake

@kushalbyatnal

Kushal Byatnal

1 month

Klarna using AI to rip out Salesforce and Workday is pretty magical at first glance.... but I've also seen this before: - company sees 7-fig Datadog bill - kicks off internal build to "save millions of dollars!" - staffs up team of eng - 6 months later, realizes their mistake

74

131

3K

2

0

25

@rahulgs

rahul

3 months

every visit to a genweb app goes straight to an llm, which renders the initial page in html all “code” is in natural language, which is “interpreted” by an llm real time user interactions are piped back into llm, which "rerenders" the page every user session is a multi-turn

Tweet media one

1

0

24

@rahulgs

rahul

1 year

next few years are going to be crazy if you're curious how many tokens are in your codebase: A bunch of our repos fit in one context window 🤯

Tweet card media

GitHub - 1rgs/token-trekker-rs

Contribute to 1rgs/token-trekker-rs development by creating an account on GitHub.

@AnthropicAI

Anthropic

1 year

Introducing 100K Context Windows! We’ve expanded Claude’s context window to 100,000 tokens of text, corresponding to around 75K words. Submit hundreds of pages of materials for Claude to digest and analyze. Conversations with Claude can go on for hours or days.

215

1K

5K

1

1

24

@rahulgs

rahul

4 years

thank you @rememberlenny , this is why we do what we do

Tweet media one

0

1

24

@rahulgs

rahul

3 years

@will_ye_ psa: this screenshot is PHOTOSHOPPED

0

0

24

@rahulgs

rahul

3 years

With Chime, we're bringing the magic of Cohere's seamless customer interaction tools to sales and marketing teams — super excited to get this out

1

0

23

@rahulgs

rahul

3 months

building a synthetic ramp this weekend with web session replay data

1

0

22

@rahulgs

rahul

4 years

🤯 @copy_ai @chris__lu @PaulYacoubian

Tweet media one

1

3

22

@rahulgs

rahul

3 months

unlike traditional AI code generation (eg copilot, chatgpt, claude artifacts, devin), which outputs code, genweb is the LLM itself llm -> code -> app ❌ llm -> app ✅ no js, no backend code - just natural language instructions and an LLM that simulates it

Tweet media one

1

0

22

@rahulgs

rahul

6 months

was able to get access without getting off the waitlist: /<owner>/<repo>?task=<description>

Tweet card media

Copilot Workspace

Copilot Workspace is a Copilot-native dev environment designed for everyday tasks.

copilot-workspace.githubnext.com

@ashtom

Thomas Dohmke

6 months

What started out as an autocomplete pair programmer is now redefining the developer experience itself. Welcome to @GitHub Copilot Workspace: The Copilot-native developer environment — a place for all to create with code instantly in natural language.

43

246

1K

1

0

22

@rahulgs

rahul

4 years

from interviewing me for my first ever job at Superhuman to writing our first check at Cohere, Vivek has been a great mentor/supporter 🙏 thank you @vsodera — wouldn't be here w/o you

@vsodera

Vivek Sodera

4 years

Proud to be one of the first investors in @CohereHQ . Their pixel-perfect screensharing experience is 🤯! If you're a head of support, customer success, QA, onboardings, or sales, and want to use Cohere at your company, DM me. /cc @yunyu_l @rahulgs @jasonhfwang

2

1

32

0

0

22

@rahulgs

rahul

3 months

genweb is a proof of concept for now, but with faster models and cheaper inference, this could soon be how all software is made software 2.0 apps are malleable and squishy, not rigid and rules-based like it is today (1) not every feature needs to be described, and the model

Tweet media one

1

0

22

@rahulgs

rahul

3 years

vc: what’s ur mrr me: haha we’re not sharing rn vc: haha how many customers do u have and what is the average deal size

4

0

21

@rahulgs

rahul

4 months

teaching llama3 to reason in "grid" with synthetic data 👀

Tweet media one

Tweet media two

Tweet media three

Tweet media four

1

0

21

@rahulgs

rahul

3 years

when other founders ask about engineers you want to hire ft @hankai1998

Tweet media one

3

0

21

@rahulgs

rahul

3 years

unintended consequence of running paid ads: everyone you’ve talked to in the last 36 years sends u a screenshot

0

0

21

@rahulgs

rahul

4 months

"LLMs are not enough" is going to age extremely poorly from @leopoldasch

Tweet media one

@mikeknoop

Mike Knoop

4 months

No one can beat the 2019 ARC-AGI benchmark. We've stalled. LLMs are not enough. Frontier research has gone closed source. We need new ideas. Maybe from you? Thrilled to announce @arcprize with @fchollet A $1,000,000 competition to beat ARC and re-start open AGI progress

34

147

782

3

2

21

@rahulgs

rahul

3 years

🔥🔥🔥 @GhorbaniAmir

Tweet media one

1

2

21

@rahulgs

rahul

3 years

Announcing my new $200 fund

Tweet media one

3

0

20

@rahulgs

rahul

2 years

Try Clarity for yourself:

An app for layered, depth-first reading — start with summaries, tap to explore details, and gain clarity on complex topics.

clarity.rahul.gs

1

1

20

@rahulgs

rahul

3 months

Microsoft’s vision was “A computer on every desk and in every home” Now it’s “an IT guy in every machine and in every home”

@cognition_labs

Cognition

@cognition_labs

3 months

Fixing a cloud virtual machine bricked by the CrowdStrike outage is an ideal job for AI agents like Devin. Nobody wants to do it, but it needs to be done — millions of times. Delegating these tasks to AI frees up engineers to do more interesting work. Here’s Devin’s fix:

39

65

427

1

0

20

@rahulgs

rahul

1 year

@bryanhpchiang this uses logits atm, which we can get from oss models, can build a simpler version that uses openai, but the main issue is multiple network round trips to openai

2

1

19

@rahulgs

rahul

19 days

scott joplin made a song about retrieval augmented generation in 1899 🙏

Tweet media one

1

0

19

@rahulgs

rahul

11 months

board trying to get @sama back

Tweet media one

@verge

The Verge

11 months

Breaking: OpenAI board in discussions with Sam Altman to return as CEO

723

2K

8K

0

0

18

@rahulgs

rahul

3 months

"By switching to the new gpt-4o-2024-08-06, developers save 50% on inputs ($2.50/1M input tokens) and 33% on outputs ($10.00/1M output tokens) compared to gpt-4o-2024-05-13." this is likely because sampling json significantly cuts down number of tokens to sample from a model a

@sama

Sam Altman

3 months

by very popular demand, structured outputs in the API:

425

703

6K

0

0

18

@rahulgs

rahul

3 years

unlike set-it-and-forget-it saas dashboards, users spend hours in Cohere every day after lots of profiling, caching, preloading and data splitting, the dashboard feels like ⚡, as all software should be big wins from (and welcome to the team!) @JustinMMott

0

0

18

@rahulgs

rahul

4 years

Been accumulating these for a while, great to get it out! cc @toddg777 @Scratchpad

@CohereHQ

Cohere

4 years

💕 Announcing the Cohere Wall of Love 💕 Thank you everyone for all your support! big launches coming next week 😉

0

1

13

1

0

18

@rahulgs

rahul

2 months

@DavidCahn6 Amazon/Anthropic

0

0

18

@rahulgs

rahul

1 year

We’re able to grid search LLM training, and then quickly spin up 100s of inference servers each with webhooks. I don’t need to think about Docker, k8s, DNS, gpu quotas, or load balancing. Everything just works. Just not possible on any other platform today Congrats team!

0

0

17

@rahulgs

rahul

3 years

Awesome to see @CohereHQ Hint: 🎥

@toddg777

Todd Goldberg

3 years

Here's a snapshot. This should make it easier for you to see the details throughout the piece 😁

Tweet media one

1

0

16

1

2

18

@rahulgs

rahul

2 years

I was wondering how many tokens are in our repos, so I asked gpt4 to write me a rust library for an added challenge, I asked it support globs, use parallelism, and add nice formatting and colors try token_trekker_rs:

Tweet media one

1

1

16

@rahulgs

rahul

3 months

HTML <--- WE ARE HERE NOW Blink Skia Draw Commands Quartz Compositor Core Graphics Core Animation Metal macOS Display Driver Darwin GPU (GPU commands, framebuffers) Display Controller (timing signals) HDMI Signal (TMDS protocol) Display Hardware (pixel matrix) Pixels

@karpathy

Andrej Karpathy

3 months

@rahulgs it's cool but your mind is still trapped within the confines of the system - <button>s, <div>s... irrelevant intermediates, blinding you from the truth. that there are no <button>s

28

16

537

2

0

17

@rahulgs

rahul

4 years

coming soon 👀

Tweet media one

1

1

17

@rahulgs

rahul

13 days

if you're not maxing out your chatgpt o1-preview weekly limit ur ngmi

3

0

17

@rahulgs

rahul

3 months

with high temperature, every visit to a genweb page results in a unique experience example running on llama3-8b on @togethercompute emoji todo list: any tasks you add becomes a set of emojis, automatically each screenshot is the same app:

Tweet media one

2

0

16

@rahulgs

rahul

4 years

joined by @Soma_Capital , @BoxGroup , @ChapterOne , @ShrugCap , and rockstar angels @zachperret , @eglyman , @karimatiyeh , @rahulvohra , @vsodera , @ericwu01 , @shravvmehtaa , @athenakan_ , @eladgil , @naval , @dwr , @nickcandito , @jaltma , @oscrhong , @varunsrin , @zcabrams , @tinab , @myprasanna +more

1

0

15

@rahulgs

rahul

4 years

woah crazy to see demos in the wild, thanks for making this @elie2222

@elie2222

Elie Steinbock

4 years

Quick demo. @CohereHQ is that easy Not affiliated

0

2

16

1

1

15

@rahulgs

rahul

1 year

openai releases json structured generation with a jsonformer-esque api

Tweet media one

0

1

14

@rahulgs

rahul

20 days

actually very good, just works

@WisprAI

Wispr Flow

21 days

Today, we’re excited to announce Wispr Flow 🚀 Just speak, and Flow writes for you, everywhere on your computer. No BS, no waitlist. Feel the magic 👉

168

182

1K

0

0

15

@rahulgs

rahul

2 years

⚙️ How does it work? Clarity uses recursive summarization to reduce long texts into digestible paragraphs. Click on any sentence to reveal the most similar sentence in the next level summary, giving you complete control over how much detail you want to explore. 🔍

1

0

14

@rahulgs

rahul

3 years

you heard it from the man himself

@packyM

Packy McCormick

3 years

Less than 24 hours in: 2.4k unique views, 1.4k job clicks, 70 applications submitted🔥 & a BANGER of a featured role: Software Engineer @CohereHQ Friend who told me about Cohere: “Cohere is the real deal. Software that will change software forever.”

4

5

46

0

0

14

@rahulgs

rahul

4 months

blind people 🤝 ai agents accessibility tags

0

1

14

@rahulgs

rahul

3 months

codegen providers have historically been forced to pick between base sota or finetuned non-sota models finetuning of sota models - 4o, llama 405b and mistral large enough will make codegen work significantly better, especially for large codebases expect a massive uptick in

2

0

14

@rahulgs

rahul

2 years

💡 Why? In a world where digital attention is getting cheaper, human attention becomes valuable. Clarity lets you read anything and get to the point faster, making the most of your precious time. 🕒

1

0

13

@rahulgs

rahul

2 years

GPT-4.5 leaked

Tweet media one

0

0

13