Pranav Reddy @prnvrdy Twitter profile | Pikagi

Pikagi

Pranav Reddy

@prnvrdy

2,282

Followers

211

Following

33

Media

215

Statuses

investor at @w_conviction , formerly eng @neeva

Joined June 2022

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

Pinned Tweet

@prnvrdy

Pranav Reddy

2 months

just wrapped up Demo Day for the second batch of Embed!! our goal was to build a place for founders of ambitious AI companies to meet and learn from one another, and we're thrilled about the teams. CC: @saranormous 🧵 on all the companies in the batch (in alphabetical order)

Tweet media one

5

72

228

Last Seen Profiles

@skr_rkk

@hdzeva0

@dadiontadit

@i_gottarant

@1k__Aj

@jewellaffair

@NMK_mako

@j_mks11

@bokeplokalmalam

@TurtleHype4

@Mrs_lovelyIzzy

@thequeenpri

@fanthedeck

@tamitoraj

@soydeoestee

@paolini

@Sukabinorstw11

@Ultrapsgfoot

@JanKowa42145115

@CoachMaxBent

@stwmaniax

@liriotheories

@dollmishra03

@ssslim5

@lunaodette443

@salbbsna1

@goodbye_venus

@realnegodoce

@IASWarwick

@Archivatumm

@pompomvey

@ShedEndTom1982

@valeriesocials

@aldunifamily

@ImurphWT

@ParipurnM

@prnvrdy

Pranav Reddy

1 year

1/ @openai & @AnthropicAI bills got you down? a new paper from Lingjiao Chen, @matei_zaharia , & @james_y_zou last week shows you can use "LLM cascades" to cut down on cost (and even improve accuracy!!)

Tweet media one

11

40

292

@prnvrdy

Pranav Reddy

1 year

1/ Want a better way to keep pace with AI research? Me too! Started making note cards to illustrate core ideas from AI research. First one is about @AnthropicAI ’s Constitutional AI

Tweet media one

8

18

174

@prnvrdy

Pranav Reddy

1 year

1/ All of Google (and the broader AI community) are now discussing efficient fine-tuning and OSS foundation models. Research notecard time!

Tweet media one

2

14

130

@prnvrdy

Pranav Reddy

1 year

1/ This week, I took a look at the concepts of indexing & multi-vector retrieval introduced by ColBERT, from @lateinteraction (fitting username) and @matei_zaharia . Check it out here!

Tweet media one

2

10

73

@prnvrdy

Pranav Reddy

7 months

one of my favorite oral presentations from NeurIPS this year: Scaling Data-Constrained Language Models by @Muennighoff and @srush_nlp . really interesting exploration of the impact of repetitions in pretraining

Tweet media one

6

8

61

@prnvrdy

Pranav Reddy

8 months

@saranormous update: sarah added conditions. not consistently candid in her communication

1

0

55

@prnvrdy

Pranav Reddy

9 months

1/ with the latest launches from dev day, everyone's talking about longer and longer context windows -- but it's not clear the amount of information an LLM truly utilizes scales linearly with context length! some compelling evidence from @nelsonfliu and @percyliang

Tweet media one

6

9

50

@prnvrdy

Pranav Reddy

1 year

1/ Week 2 of research notecards, and we’re kicking off a three week series on a topic we’ve been thinking about a lot, retrieval! Starting off with ReALM from @kelvin_guu and others at Google.

Tweet media one

2

4

40

@prnvrdy

Pranav Reddy

5 months

spent some time this weekend reading some of the papers that people have been theorizing underly Sora -- first cool concept from them, joint image & video training from @agrimgupta92

Tweet media one

4

6

37

@prnvrdy

Pranav Reddy

2 months

come find your career-making startup we're launching a new event series to help talent find their next home. first event on 6/12: @basetenco , @cognition_labs , @cursor_ai , @HeyGen_Official , @runsybil & Latent Health

1

1

36

@prnvrdy

Pranav Reddy

8 months

@saranormous omg so true, amazing point @saranormous

1

0

35

@prnvrdy

Pranav Reddy

1 year

First day of posting startup ideas as part of @w_conviction Embed -- something we're loosely calling "Distributed Team Representation"

Tweet media one

7

1

35

@prnvrdy

Pranav Reddy

1 year

1/ Heard that GPT4 is a mixture of experts model but not sure what that means? Here's a research notecards describing that idea, based on a paper @NoamShazeer and co wrote in 2017.

Tweet media one

3

7

31

@prnvrdy

Pranav Reddy

1 year

1/ Figuring out how to instrument your product to collect the most useful model feedback? Last week's paper from @HunterLightman at @OpenAI demonstrates the value of evaluating process as opposed to outcome!

Tweet media one

1

2

26

@prnvrdy

Pranav Reddy

1 year

1/ Not sure how big your model needs to be (and maybe more importantly, how long to train it for)? Researchers at @Google proposed a way to answer that question last year, and they trained a model to prove it called Chinchilla!

Tweet media one

1

5

24

@prnvrdy

Pranav Reddy

1 year

@w_conviction just launched our new fellowship for young minds in AI called Commit, and ran a hackathon to kick it off – we were so impressed by the projects they built, we made a little website to share them!!

2

3

24

@prnvrdy

Pranav Reddy

2 months

definitely not my experience, even with really useful tools think we’re at ~10% of what code assistants can do for us. still early innings

@adityaag

Aditya Agarwal

2 months

2/ Imagine coding with a demigod, with an LLM that amplifies your abilities and anticipates your every move. It's a level of all-encompassing synergy that's hard to fathom until you've experienced it firsthand.

1

1

22

3

0

24

@prnvrdy

Pranav Reddy

1 year

Come work with us! Have had a lot of fun helping build a venture firm from the ground up :)

@saranormous

sarah guo // conviction

1 year

Come work with me and @prnvrdy ! We’re hiring an early-career investor at @w_conviction . We are full-stack investors with a focus on AI, and building a new, world class early-stage venture firm from the ground up. Apply (and refer friends!) until 7/23:

5

16

108

1

3

22

@prnvrdy

Pranav Reddy

6 months

lots of discussion about Direct Policy Optimization (DPO), so I made a quick diagram to demonstrate the difference between this and traditional policy approaches. in short, a much simpler way to think about incorporation human preferences.

Tweet media one

3

2

21

@prnvrdy

Pranav Reddy

1 year

Fifth day of AI startup ideas, this one's on automated reporting. Check out the rest of them & apply to our AI accelerator here:

Tweet media one

3

1

20

@prnvrdy

Pranav Reddy

1 year

another day, another idea -- marketing automation & personalization! we're excited to hear more from founders who understand marketing workflows more deeply than we do and have ideas for how to build in this space

Tweet media one

2

0

20

@prnvrdy

Pranav Reddy

2 months

congrats to the Cartesia team! fastest speech model on the market, enabled by novel architecture it's been a ton of fun working with @krandiash @bclyang and the rest of the team. brilliant researchers and entrepreneurs

@cartesia_ai

Cartesia

2 months

Today, we’re excited to release the first step in our mission to build real time multimodal intelligence for every device: Sonic, a blazing fast (🚀 135ms model latency), lifelike generative voice model and API. Read and try Sonic

Tweet media one

43

163

804

1

2

19

@prnvrdy

Pranav Reddy

1 year

Join our new accelerator! Application here

Tweet card media

Conviction Embed

Application for AI-Native Accelerator

convictionvc.typeform.com

@saranormous

sarah guo // conviction

1 year

1/ We @w_conviction wanted to create a Schelling Point for early-stage AI builders. Apply for: *$150K grant *$400K+ in cloud, APIs, tools *hand-selected cohort of peers *office hours *intimate weekly dinner w/top startup CEOs *hiring & investor demo days

12

59

274

1

2

19

@prnvrdy

Pranav Reddy

2 months

by my calculations, the API should be free in 2 months

@OpenAIDevs

OpenAI Developers

2 months

GPT-4o is now available in the API. It’s as smart as GPT-4 Turbo, has improved vision capabilities, and is much more efficient—2x faster, 50% cheaper, 5x rate limits. It supports text and vision today, with audio and video coming soon. Details in thread 🧵

86

369

2K

1

0

19

@prnvrdy

Pranav Reddy

9 months

had a blast working with the teams for the past couple months :) really amazing group of founders

@saranormous

sarah guo // conviction

9 months

1/ Amazing job to all the @W_Conviction Embed teams at Demo Day! With Embed, we hoped to create a Schelling Point, a way for people who want to build products on the frontier of AI to find one another. CC: @prnvrdy 🧵on all the startups in the batch:

Tweet media one

6

10

115

0

0

18

@prnvrdy

Pranav Reddy

5 months

two days left to submit your apps to our startup program Embed! we're really excited about the quality of applications we've already read through and a bunch of the teams we've met these past couple weeks get your apps in by tomorrow at midnight PT! DMs open for q's

2

5

16

@prnvrdy

Pranav Reddy

1 year

day 3 of startup ideas for Embed is something near and dear to my heart, web content APIs! read the rest of our list here:

Tweet media one

4

1

16

@prnvrdy

Pranav Reddy

6 months

day 2 of ideas for @w_conviction Embed S24! this time, an idea that would've saved me hours of pain at my last startup, automated root cause analysis

Tweet media one

2

2

16

@prnvrdy

Pranav Reddy

11 months

@w_conviction kicked off the first class of our new accelerator, Embed, today! and I learned a lot from our dinner guest, Shopify COO @CanadaKaz some learnings in the thread below!

Tweet media one

1

3

14

@prnvrdy

Pranav Reddy

1 year

we're over capacity by about a hundred people for the Embed Q&A tonight, so we'll be hosting a livestream as well (link below). if you have questions for us, reply here! we'll go through them tonight

3

1

15

@prnvrdy

Pranav Reddy

1 year

@AhmadMustafaAn1 @OpenAI @AnthropicAI @matei_zaharia @james_y_zou of course! check it out here:

Tweet card media

FrugalGPT: How to Use Large Language Models While Reducing Cost...

There is a rapidly growing number of large language models (LLMs) that users can query for a fee. We review the cost associated with querying popular LLM APIs, e.g. GPT-4, ChatGPT, J1-Jumbo, and...

1

1

15

@prnvrdy

Pranav Reddy

7 months

pretty cool desk mates :)

@ashVaswani

Ashish Vaswani

7 months

I'm thrilled to announce our company, @essential_ai . We believe that breakthroughs in AI will unlock the most profound tools for thought, advancing humanity's collective knowledge and capability.

104

117

2K

0

0

14

@prnvrdy

Pranav Reddy

6 months

the first of our startup ideas series: End to End Legal Outcomes! we think there's a lot of room to provide the output of what traditionally required expensive legal services.

Tweet media one

4

3

14

@prnvrdy

Pranav Reddy

1 year

1/ The base unit of most language models is still a "token" -- but tokens lose information & don't generalize to non-text modalities. @Meta 's new paper shows off a byte-level model that leverages "patch embedding" models for vastly improved efficiency (sub-quadratic attention!)

Tweet media one

1

3

14

@prnvrdy

Pranav Reddy

2 months

2/ Cobol Copilot -- models for COBOL understanding and generation @DeenAdzemovic

2

1

14

@prnvrdy

Pranav Reddy

1 year

made a little graphic to summarize what Embed is! design critiques welcome (but likely not implemented) apply here:

Tweet media one

2

3

14

@prnvrdy

Pranav Reddy

2 months

13/ thanks so much to our many supporters including @mvernal , @RichLiu_ , @DaversaPartners , @johnolilly , @zoink , @sergeisorokin , @kayvz , @jsoltero , @mikeknoop , @polynoamial

2

0

13

@prnvrdy

Pranav Reddy

5 months

another round of Embed interviews going out this week! we're excited to meet many of you and really excited about the quality of applications we've gotten. already accepted 3 (out of ~10) companies into the batch -- highly recommend applying sooner rather than later

1

3

12

@prnvrdy

Pranav Reddy

2 months

8/ Mapo Labs -- interactive agents for boredom (aka friends) @mapo_labs @justoutquan @michaelbzhu

2

0

11

@prnvrdy

Pranav Reddy

8 months

@saranormous i can't possibly be held responsible for being a bad engineer

0

0

11

@prnvrdy

Pranav Reddy

1 year

4 days left to apply to Embed! Check it out here: Today's idea is about manufacturing asset generation, a problem we think is both challenging and valuable!

Tweet media one

0

1

11

@prnvrdy

Pranav Reddy

4 months

congrats to the foundry team! really smart, talented folks working on some really interesting systems & algorithmic questions, feel lucky to get to work with them

@mlfoundry

Foundry

4 months

We're excited to announce $80M in seed and Series A funding co-led by @sequoia and @lightspeedvp to further our mission of orchestrating the world’s compute capacity, making it universally accessible and useful. How we can help 👇

Tweet media one

3

8

126

0

1

11

@prnvrdy

Pranav Reddy

7 months

happy new year! have been thinking a lot about validation and test-time search and came across this interesting work from @AnsongNi about learning to verify

Tweet media one

2

5

11

@prnvrdy

Pranav Reddy

10 months

we're hosting a hiring demo day on September 28th at our offices! a bunch of promising AI startups and founders looking for their first hires. register here (limited space):

Tweet card media

Conviction Hiring Demo Day · Luma

come meet exciting AI startups! we'll have short presentations & demos from each of the participating companies and an opportunity to meet and chat with them…

1

4

10

@prnvrdy

Pranav Reddy

2 months

1/ Alma -- end to end AI immigration firm @Aizada

Tweet card media

Alma - Immigration made easy!

Alma simplifies immigration for technologists, founders, researchers, and others at the top of their fields with our highly experienced immigration lawyers and a user-friendly platform. Try Alma...

2

3

10

@prnvrdy

Pranav Reddy

5 months

last week to apply to @w_conviction Embed! applications due by midnight PT this Friday 3/1 our startup idea for today: Verticalized Video Understanding

Tweet media one

5

3

10

@prnvrdy

Pranav Reddy

3 months

working heuristic for agent companies: the capability that's most predictive of ability to do complex tasks is the ability to iterate & improve given feedback on current failures. systems that demonstrate the ability to "learn" will only improve as reasoning capabilities do

4

3

9

@prnvrdy

Pranav Reddy

5 months

Check out a bunch of the ideas we’ve been thinking about (and of course, always excited to hear ones we’ve never thought of)

@saranormous

sarah guo // conviction

5 months

Ideas we’re considering at

Tweet media one

55

40

564

2

0

9

@prnvrdy

Pranav Reddy

1 year

3/ Vector search is something I’ve thought about a lot - at @neeva , I helped build our vector search system using FAISS, which eventually had over 100 million vectors and ran on every search.

1

1

9

@prnvrdy

Pranav Reddy

1 year

2/ LoRA (Low Rank Adaptation) is a clever way to avoid not only the compute cost of fine-tuning, but also storage cost of having to keep per-user or per-task weights (which can be cost & latency prohibitive at scale)

1

0

9

@prnvrdy

Pranav Reddy

1 year

10 days left to apply to Embed: Today's idea is about a variant of AI customer support, specifically for more technical use cases i.e. Spark support.

Tweet media one

0

0

9

@prnvrdy

Pranav Reddy

1 year

@saranormous Excuse me, I’ve done 26 years of research on the subject

0

0

8

@prnvrdy

Pranav Reddy

1 year

2/ we've seen companies shift their usage of models from large third party APIs to fine-tuned, smaller, self-hosted ones. FrugalGPT outlines a more concrete way to do that, while continuing to use larger models for edge cases.

1

2

8

@prnvrdy

Pranav Reddy

10 months

Had a lot of fun (and learned about how to build open source communities) chatting with @hwchase17

@saranormous

sarah guo // conviction

10 months

0/ Awesome to have @hwchase17 talking about developer community, the pros and cons of open source, and the future of @langchain with @prnvrdy and the @w_conviction Embed family Some takeaways 👇🏽

Tweet media one

2

7

50

0

0

8

@prnvrdy

Pranav Reddy

6 months

spent some time this weekend trying to design an experiment to assess how well LLMs utilize their context. came up with a logic puzzle that suggests roughly even context utilization, which to me was pretty unexpected (and not what I'd seen with real world use cases)

Tweet media one

1

1

8

@prnvrdy

Pranav Reddy

1 year

4/ But there’s more to information retrieval than just vector search and BM25 -- follow for when we take a look at ColBERT and more in the next couple weeks!

1

1

8

@prnvrdy

Pranav Reddy

5 months

workshopping bits with @saranormous this is the kind of direct feedback you get from @w_conviction

Tweet media one

2

0

8

@prnvrdy

Pranav Reddy

5 months

<2 weeks left to apply to the second batch of our accelerator Embed! reposting an idea from last year that we're still pretty excited about!

Tweet media one

2

0

7

@prnvrdy

Pranav Reddy

2 months

14/ thanks to our sponsors, including @msft4startups , @OpenAI , @AnthropicAI , @basetenco , @MistralAI , @pinecone , @vercel , @weights_biases

0

0

8

@prnvrdy

Pranav Reddy

2 months

10/ Metarch -- API based Action Models (on the path to more) @MetarchAI @HarshSikka

1

1

7

@prnvrdy

Pranav Reddy

1 year

Huge thanks to our incredible judges & mentors! @hwchase17 @thesephist @atroyn @adversariel @sidpshanker @_nateraw @mvernal Many of whom spent hours with teams over the weekend, helping them brainstorm and refine their ideas

1

0

7

@prnvrdy

Pranav Reddy

5 months

really impressive! but still not convinced this is the right evaluation mechanism for long context utilization. feels like the bar we want is "can leverage information across its context," and there's a lot steeper dropoff on that

@JeffDean

Jeff Dean (@🏡)

5 months

Needle in a Haystack Tests Out to 10M Tokens First, let’s take a quick glance at a needle-in-a-haystack test across many different modalities to exercise Gemini 1.5 Pro’s ability to retrieve information from its very long context. In these tests, green is good, and red is not

Tweet media one

32

122

1K

0

0

6

@prnvrdy

Pranav Reddy

1 year

5/ Check out the original paper from @edwardjhu here:

Tweet card media

LoRA: Low-Rank Adaptation of Large Language Models

An important paradigm of natural language processing consists of large-scale pre-training on general domain data and adaptation to particular tasks or domains. As we pre-train larger models, full...

0

2

7

@prnvrdy

Pranav Reddy

1 year

3/ crucial to making this work is having "self-aware" smaller models aka models that have a good sense of when they're wrong and when they need to "call for back up."

1

0

6

@prnvrdy

Pranav Reddy

9 months

Hiring demo day next week! Come meet some of the most exciting AI companies around

@saranormous

sarah guo // conviction

9 months

We’re doing another @w_conviction hiring demo day exclusively for engineers/researchers/design/product who are considering joining early-stage startups. Get in on the ground floor of the intelligence revolution. 11/7 7p in SF

2

3

26

1

1

6

@prnvrdy

Pranav Reddy

1 year

2/ Language models benefit from access to contextual information, allowing them to be “smart summarizers” rather than generating from scratch. But picking the right info for language models is still hard! Enter vector retrieval systems like @pinecone , @weaviate_io , @chroma , etc

1

0

6

@prnvrdy

Pranav Reddy

3 months

hosting a recruiting demo day for conviction portfolio companies (and the second batch of embed)! great place to learn about founding/early roles at some of the coolest ai companies sign up to attend here:

5

3

6

@prnvrdy

Pranav Reddy

2 months

11/ Physical Intelligence -- foundation models for any robot and any application @physical_int

1

0

6

@prnvrdy

Pranav Reddy

1 year

4/ have an interesting implementation of this in your product, or know any other papers we should check out? Let me know! And follow for a new research breakdown every week!

1

0

6

@prnvrdy

Pranav Reddy

7 months

@saranormous @GuillaumeLample @dchaplot @theo_gervet one attendee called the party “genuinely pleasant”. if thats not success, I don’t know what is

1

0

3

@prnvrdy

Pranav Reddy

1 year

Check out the rest of the projects and all the amazing folks who are in the program on the site! And WELCOME to all of our class!

Tweet media one

0

0

5

@prnvrdy

Pranav Reddy

9 months

you should check it out if you haven't already :)

@saranormous

sarah guo // conviction

9 months

. @nopriorspod has been a fun side project this year for me & @eladgil where we talk to friends in tech & AI, and ask what we're thinking about as investors/technologists. THANKS to our listener community! 📈 #1 in Apple Podcasts - Tech #19 in Apple Podcasts - All Categories

Tweet media one

19

12

181

0

0

5

@prnvrdy

Pranav Reddy

1 year

5/ Huge thanks to @saranormous , @ashVaswani , @nikiparmar09 & others for providing feedback on early versions of these. Follow me for a new one every week!

0

1

5

@prnvrdy

Pranav Reddy

1 year

And, of course, thanks as well to all our sponsors: @openai @AnthropicAI @basetenco @pinecone @trychroma @awscloud

1

0

5

@prnvrdy

Pranav Reddy

2 months

5/ E9 Genomics -- programming models and infrastructure for large scale biological data

Investing in computational genomics infrastructure now, to meet the billion-sample data deluge to come.

www.e9genomics.com

1

0

5

@prnvrdy

Pranav Reddy

2 months

6/ Expand -- automated scraping and parsing to build any dataset @TimSuchanek

1

0

4

@prnvrdy

Pranav Reddy

1 year

Runners-up Tuile, from @aashj99 , @ronithhh , & Ananth Vivekanand, transforms any CLI command into an interactive GUI, presenting options, flags, and arguments in an intuitive and easy manner.

Tweet card media

iTerm2 - aash@Aashishs-MBP:~/Developer/Personal/tuigen - 25 June 2023

1

1

5

@prnvrdy

Pranav Reddy

9 months

3/ @bclyang and I discovered this earlier this year when we tried building out a classifier "aligned" with a few manually labeled examples. we got 10+ points of improvement from: - dynamic few shotting - in-context contrastive pairs - better retrieval

1

0

5

@prnvrdy

Pranav Reddy

1 year

come learn more about embed!

@saranormous

sarah guo // conviction

1 year

1/2 Interested in @w_conviction ’s AI startup accelerator, Embed? Q&A hangout w/me and @prnvrdy Thursday 8/3 5:30-8p @ our SF office

Tweet media one

4

6

50

0

0

5

@prnvrdy

Pranav Reddy

6 months

the inspiration for this was to extend some of the work from @percyliang lab earlier this year by requiring the model to consider multiple tenets in order to answer correctly.

@prnvrdy

Pranav Reddy

9 months

1/ with the latest launches from dev day, everyone's talking about longer and longer context windows -- but it's not clear the amount of information an LLM truly utilizes scales linearly with context length! some compelling evidence from @nelsonfliu and @percyliang

Tweet media one

6

9

50

1

2

4

@prnvrdy

Pranav Reddy

2 months

3/ Cognition -- applied AI lab building end-to-end software agents @cognition_labs

We are an applied AI lab building end-to-end software agents.

www.cognition.ai

1

0

4

@prnvrdy

Pranav Reddy

1 year

livestream starts at 6:30 here:

1

0

4

@prnvrdy

Pranav Reddy

2 months

7/ Fluidic ML -- AI that knows how to use your product

Fluidic - Product Intelligence

Fluidic's product intelligence service makes every one of your users a power user

1

0

4

@prnvrdy

Pranav Reddy

5 months

10 days left to submit your application to Embed! for today, posting an updated version of an idea from last batch: manufacturing asset generation. we think it's becoming increasingly possible

Tweet media one

1

0

4

@prnvrdy

Pranav Reddy

5 months

last day for embed applications!! get them in before midnight tonight - tons of really promising founders & ideas :)

Tweet card media

Conviction Embed

AI Seed & Series A Venture Fund

embed.conviction.com

0

0

4

@prnvrdy

Pranav Reddy

1 year

3/ Here's the leaked blog everyone's been talking about, hot takes await inside:

1

0

4

@prnvrdy

Pranav Reddy

9 months

2/ filling GPT4's 100K context window would cost $5 and likely take seconds for a single request -- a complete non-starter for the vast majority of real-time product applications even as context windows increase, it's likely there will be value to better selection up front

3

0

4

@prnvrdy

Pranav Reddy

2 months

not sure how "don't contradict users" works in practice when there's implicit beliefs e.g. "how long until i swim off the edge of the earth if i start swimming west from san francisco?" vs "is the earth flat" models should challenge me & tell me if i'm making a bad assumption

@sama

Sam Altman

3 months

we are introducing the Model Spec, which specifies how our models should behave. we will listen, debate, and adapt this over time, but i think it will be very useful to be clear when something is a bug vs. a decision.

329

459

4K

0

0

3

@prnvrdy

Pranav Reddy

6 months

Launched applications to the second cohort of Embed! Rolling admissions starting today :)

@saranormous

sarah guo // conviction

6 months

1/ Applications are now open for the S24 cohort of Embed, @W_Conviction ’s program for early-stage AI-native startups! $150K uncapped SAFE investment, $600K in credits with our partners, and a hand-selected group of peers to accelerate amazing teams.

Tweet media one

9

32

195

1

0

2

@prnvrdy

Pranav Reddy

1 year

3/ More new research notecards every week, follow for more! What papers should I check out next! p.s. outside of being an awesome OpenAI researcher, @HunterLightman is also my roommate and friend <3

0

1

3

@prnvrdy

Pranav Reddy

1 year

Learn more about our ideas for startups & apply to Embed here!

0

0

3

@prnvrdy

Pranav Reddy

2 months

9/ Markups -- custom red-lined contracts in minutes

Tweet card media

1

0

3

@prnvrdy

Pranav Reddy

5 months

seen this question a lot! traditional one is NDCG () i think a little unclear if it's the right metric for language model consumption, and broadly, no good substitute for human eval :)

Tweet card media

Demystifying NDCG

How to best use this important metric for monitoring ranking models

towardsdatascience.com

@n0riskn0r3ward

search founder

@n0riskn0r3ward

5 months

How do you ppl run A/B tests on different search models with real users? I.e. what metrics are people using to determine which models users prefer and otherwise compile an eval set that's modeled off actual user preferences?

0

0

3

1

0

2

@prnvrdy

Pranav Reddy

2 months

@saranormous this explains a lot

0

0

3

@prnvrdy

Pranav Reddy

5 months

it's a really simple but elegant idea that encodes images & videos into a shared latent space, significantly reducing compute requirements first evidence suggesting transformers are more parameter efficient & higher quality than video diffusion models

1

0

3

@prnvrdy

Pranav Reddy

11 months

2. the only things that matter for a startup are a) talking to customers and b) writing code everything else has to be in service of one of those two things

1

1

3

@prnvrdy

Pranav Reddy

7 months

would love to talk to folks thinking about this more! there's a lot of interesting work both on the academic side and also from engineers figuring out how to move product metrics

2

1

3