Ivan Leo @ivanleomk Twitter profile | Pikagi

Pikagi

Ivan Leo

@ivanleomk

2,293

Followers

900

Following

408

Media

3,021

Statuses

I work on applied AI projects and spend my spare time tinkering with models for fun. I write longer tweets at

Singapore

https://t.co/02byZNKMa8

Joined March 2011

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

Pinned Tweet

@ivanleomk

Ivan Leo

26 days

Using whisper is so 2023. Just use gemini, pass in the raw audio directly and prompt the model directly with all the questions you have. With instructor, we can get - The exact mispronounced word - The timestamp when we did it - Advice on how to do better Flash truly is the

1

0

24

Last Seen Profiles

@GhanshyamB11592

@misschloecox

@themostchic

@Shethamm

@SandeepT0100

@widictya

@rasiwarindra

@kenyanicon001

@RobodebtRC

@LopicicMilos

@IBD_Alexis

@CatlogShop

@Rk05265656

@mojocoffeenz

@penyukastw21

@yeahbrizzy

@stwmaniax

@mymoonaaah

@samheckle

@KhushiSharma97

@Beleza1965Pura

@GainesCorwin

@RAYSAN711

@cjmarsh12

@wowimgutted

@Wes489076910805

@Shravankumardp

@cr_etina

@misaki_yakumo08

@bokeplokalmalam

@Reo5419233

@serabipemersatu

@PessAsunci

@thadynastyy

@Airdrop_Crakers

@cre8tions

@ivanleomk

Ivan Leo

9 months

I've had some people ask me for advice on ML so I compiled some of my thoughts into 3 tips that I wish I had thought of/done more of when I first started I'm honestly still a beginner but I figured I'd offer some thoughts as to how someone can juggle this with a fulltime job

9

34

280

@ivanleomk

Ivan Leo

11 months

Literally got into machine learning because of @jeremyphoward 's FastAI course a few months ago. This is surreal. This just made my entire week.

Tweet media one

9

4

182

@ivanleomk

Ivan Leo

3 months

Save 50% on your OpenAI Bill ( even on fine-tuned models) with batch jobs using Instructor Just use our new BatchJobs object :)

1

3

109

@ivanleomk

Ivan Leo

5 months

@NielsHoven This was my experience to some degree. I was using algebra to solve problems in the 5th grade but was penalised due to the education system requiring people to use models for math. I do think I would have discovered how much more fun math was if I had been able to explore/learn

3

2

82

@ivanleomk

Ivan Leo

11 months

We had some good results before but with some careful data selection, our fine tuned GPT3.5 models managed to generate summaries an order of magnitude faster and better than GPT-4. Read here :

Tweet media one

7

6

81

@ivanleomk

Ivan Leo

5 months

Officially unemployed today - next step Bangkok for some Muay Thai and much needed rest before starting full time with @jxnlco soon. Time to speed run some distributed Rust for fun

7

2

80

@ivanleomk

Ivan Leo

1 month

1/ If you're building a RAG application, these problems probably sound familiar: 1. Irrelevant search results 2. Insufficient Data to create a database index 3. Multiple data sources that are out of sync 4. Untested LLM agents How do these problems manifest?

3

7

71

@ivanleomk

Ivan Leo

10 months

tl;dr : Do give @magicpatterns a try if you've got some time. It's pretty handy. I managed to generate the following UI in about 1-2 hours of work using their product including setting up all the dependencies from scratch with @shadcn 's ui package. The design itself is based off

Tweet media one

5

4

65

@ivanleomk

Ivan Leo

9 months

I used >70 A10Gs for 4 hours on demand for batch jobs on @modal_labs insane to see that I only spent < 100 bucks lol.

3

3

65

@ivanleomk

Ivan Leo

7 months

If you’re using colab to prototype, I highly suggest using a @modal_labs instance instead. You’ll pay per second and get much better GPUs and performance. I was able to download all my Nvidia libraries within around 10s yesterday. Colab took almost 10 minutes.

5

3

56

@ivanleomk

Ivan Leo

20 days

Spent some time today working on a data annotation tool to clean some web data after seeing @HamelHusain 's AI Engineer talk on data annotation tools. Took me ~10 mins to spin up with O1 and cursor and now I've annotated around 50 different chunks to be used for evaluation.

Tweet media one

2

4

53

@ivanleomk

Ivan Leo

10 months

Using @modal_labs for batch jobs is like damn I found a weird bottleneck, time to just use .map and 50 GPUs to get a 500x speed up lmfao

1

4

52

@ivanleomk

Ivan Leo

6 months

@culturaltutor Singapore would not be where it is without Air conditioning. Lee Kuan Yew famously said that one of the greatest inventions of the 21st century was the air conditioning unit haha

2

0

53

@ivanleomk

Ivan Leo

6 months

Can't wait to get started! We've got so many amazing projects in the pipeline in store for everyone :)

@jxnlco

jason liu

6 months

I'm officially boutique consulting firm now that @ivanleomk is joining me full time in a month!

7

1

92

8

1

51

@ivanleomk

Ivan Leo

6 months

Integrating Logfire with Instructor is dead simple. With just these 4 lines of code, you'll get full transparency into your entire application - from parsing errors of individual requests to the latency of individual functions down to the millisecond

Tweet media one

7

5

49

@ivanleomk

Ivan Leo

7 months

Reading the @deepseek_ai moe paper and it's really interesting how they decided to implement shared experts which are always used. Would love to see a comparison of these shared experts/MOE using the token visualisation that Mixtral did

Tweet media one

1

6

48

@ivanleomk

Ivan Leo

1 month

Enjoy up to 90% cost savings with Anthropic's new beta prompt caching feature. It's still in beta at the moment but using it in instructor takes just a few steps. First, let's initialise an Instructor instance

Tweet media one

3

3

48

@ivanleomk

Ivan Leo

6 months

Just ran 110 experiments with different hyper-parameters to find the best ones for my model with @modal_labs . All it took was 20 bucks, 40 A10Gs and containers. god bless

3

3

46

@ivanleomk

Ivan Leo

28 days

I get asked every now and then about whether to use @LangChainAI or instructor My general advice normally boils down to 1. How familiar are you with either library 2. How long are you maintaining this code for This is kind of how I've been thinking about it.

4

2

44

@ivanleomk

Ivan Leo

4 months

My first Agent!

Tweet media one

3

1

44

@ivanleomk

Ivan Leo

10 months

Constraining LLMs to use APIs without hallucinating is hard. json mode is a tad bit unreliable and offering a $200 tip to gpt isn't feasible in this economy with inflation. Instead, just define a @pydantic class and force the model to output the parameters you want using

Tweet media one

4

3

42

@ivanleomk

Ivan Leo

6 months

If you're using FastAPI and Instructor, Logfire is a perfect fit. We've just published a new guide with complex examples to show how to 1. Use streaming to consume extracted objects faster 2. Take advantage of asyncio to run multiple instructor calls in parallel

Tweet media one

1

3

38

@ivanleomk

Ivan Leo

8 months

@allgarbled I think it takes a certain level of courage to post about your learning process because it means opening it up to the world to critique and see. Much easier to just criticise without adding anything of value.

0

2

37

@ivanleomk

Ivan Leo

3 months

I am now a newfound convert of XML for prompting

10

1

37

@ivanleomk

Ivan Leo

5 months

Finished up more of the rust exercises in @algo_luca 's 100 exercises of rust. I'm pretty pumped to finally understand what a trait in rust is. Just three more modules and then I start on Rust Atomics and Locks recommended by @diptanu . Instructor-rs coming soon maybe?

Tweet media one

2

1

36

@ivanleomk

Ivan Leo

3 months

Playing with @cohere 's new Structured Extraction mode and it's significantly faster to use for structured extraction. Seems to be ~40% faster than using their normal chat completion and generating raw JSON

3

1

36

@ivanleomk

Ivan Leo

1 month

Want to give O1 a try over the weekend? Just use instructor's new JSON_O1 mode and you're good. Same old familiar api with a single loc to change

Tweet media one

3

3

33

@ivanleomk

Ivan Leo

3 months

@netcapgirl Maybe the tokenizer at fault here? Seems to lump 11 and 9 in single digits at the end

Tweet media one

5

1

31

@ivanleomk

Ivan Leo

2 years

Playing around with the new @supabase auth helpers and pleasantly suprised at how fast it was to set everything up - magic link auth in literally 10 minutes. That was including the time to provision the db god damn.

3

3

30

@ivanleomk

Ivan Leo

13 days

TIL that you should not be relying on your embeddings as a hash Embedding the same sentence 200 times gives 8 different embeddings. Cosine similarity is at least 0.9993 but imagine this inconsistency multiplied across 1000s of embeddings lol. Use FTS peeps

Tweet media one

3

2

31

@ivanleomk

Ivan Leo

6 months

Walking on a treadmill to burn calories and need some ML content? Here's a running thread which I'll keep up of things I liked that I've watched. Would appreciate any other recommendations too

3

4

30

@ivanleomk

Ivan Leo

9 months

1/ Show your work much earlier that you're comfortable with à la @swyx 's Learn In Public I'm a big fan of volume based outputs ever since @jxnlco introduced it to me and I think sharing my work in public has been the biggest difference for me. Early 2023 I was doing so but

4

2

30

@ivanleomk

Ivan Leo

9 months

2/ Start Top Down As @jeremyphoward says in Fast AI, you don't teach baseball by learning aerodynamics ( paraphrasing a bit here ) I think you shouldn't start looking into ML by finishing the entire linear algebra course from MIT and then learning pytorch before writing/using

1

4

30

@ivanleomk

Ivan Leo

9 months

Managed to implement a small scraper with GPT! I couldn't find a good html to markdown parser with python and had some edge cases with the html I was working with. Here's how I did it - first I defined a data model using @pydantic for the final result I wanted.

Tweet media one

1

2

29

@ivanleomk

Ivan Leo

1 month

Curious about how instructor's response_model parameter works under the good? I wrote up a short blog post walking you through a high level overview of how we convert your pydantic model into a validated response. Link :

Tweet card media

How does Instructor work?

How your request goes from chat completion to validated Pydantic model

www.ivanleo.com

1

1

29

@ivanleomk

Ivan Leo

11 months

Forked out @abacaj 's replit inference script to create a small Copilot with CPU! These autocompletions run off the older @Replit v1-3b model and directly hook into VScode using a small extension I coded up. Next step is to upgrade context and figure out webviews

4

1

28

@ivanleomk

Ivan Leo

9 months

Woo! The article I worked on all December is finally out. This was an insanely fun one to write with @jxnlco running these massive batch jobs at scale.

@jxnlco

jason liu

9 months

We were able to embed all of Wikipedia in < 15 minutes on Modal Labs for ~ 17$. Here's what this actually means for an organization: 1. Unrestricted Rates: Eliminates the bottleneck of rate limits for large-scale operations. 2. Rapid Experimentation: Allows quick iterations of

24

115

1K

3

2

28

@ivanleomk

Ivan Leo

1 month

RAG doesn't have to be complex. 1. @TimescaleDB for embedding search and metadata filtering 2. @Pydantic w instructor to test your LLMs 3. LLMs to generate metadata at scale Curious to find out how? @jxnlco , @avthars and I wrote about how to do so in

Tweet card media

RAG Is More Than Just Vector Search

Go beyond semantic search. Learn how to improve your RAG system with Text to SQL, filtered search, structured extraction, and evaluation driven development.

www.timescale.com

3

3

27

@ivanleomk

Ivan Leo

9 months

@visakanv Actually very interesting point! Can't seem to find an exact source but I remember reading that in Japan for example, ox plows stopped being used for a while because human labor was cheaper. Innovations take flight when they're good but also when conditions are right.

1

0

25

@ivanleomk

Ivan Leo

11 months

Organising a small drinks session for the llm/m folks in singapore 8 december. Come through for some fun conversations, sharings and a good time

Tweet card media

ML, LLMs, AI, and beer · Luma

Meetup for folks interested in machine learning, LLMs, and all things genAI with craft beers and good vibes downtown 🍻

2

5

26

@ivanleomk

Ivan Leo

9 months

Used Open AI's new large model and I keep getting rate limited, i've spent 3 hours trying to slowly pace out my requests and it keeps throwing an error. @cohere on the other hand, worked like a charm on the first try! Really highly recommend

@ivanleomk

Ivan Leo

9 months

@cohere ’s new embed job api is very good actually, prob around 5-10x faster than using their previous API It took a bunch of sentences that I previously spent almost 45 minutes embedding and returned the final product in just under 5-10 minutes.

1

2

16

4

2

26

@ivanleomk

Ivan Leo

6 months

Spent some time today working on the @elicitorg machine learning list. Checked off - 3Blue1Brown videos on Deep Learning - Karpathy's 2 hour video on micro-grad - The short introduction article on deep reinforcement learning Good productive saturday, now time to get back to

5

2

25

@ivanleomk

Ivan Leo

9 months

Moderated a small discussion yesterday in our @latentspacepod paper club by @swyx on the hugging face Mixture Of Experts article with @eugeneyan and we took some notes for the first time! Thought I'd highlight some interesting things that stood out to me (1/n)

1

10

26

@ivanleomk

Ivan Leo

11 months

Spent some time in the evening playing around with @modal_labs 's new volume feature and it's definitely a huge step up. On @TheBlokeAI 's Open hermes model, I'm getting a speedup of around 10% in terms of raw inference speed for both Cold and Warm endpoints. Volumes also tend

Tweet media one

Tweet media two

2

2

25

@ivanleomk

Ivan Leo

1 year

Finally got around to trying @swyx 's smol-menubar and its pretty solid. Worked out of the box for me and gave some good responses. Amazing to query 4 diff LLMs at the same time and get a response to your query.

1

3

25

@ivanleomk

Ivan Leo

9 months

Instead, work with a small problem in mind and then go backwards from that to learn what you need. Eg. Start Playing with hosted APIs, then use @modal_labs to see what u can do with OSS models, then go through @karpathy 's videos to learn more about the basics.

1

2

24

@ivanleomk

Ivan Leo

9 months

I'll add an additional tip here - find your people. It takes effort but I think it's worth it. I've been lucky enough to find friends like @aimuggle , @SriniN123 and @hrishioa who are equally interested in this space and doing great work. Twitter helps a lot. Discord groups are

2

2

22

@ivanleomk

Ivan Leo

9 months

Thought it was pretty cool that @LumaLabsAI promoted their internship and was like yeah join us we got 3200 A100 GPUs

Tweet media one

1

3

22

@ivanleomk

Ivan Leo

1 year

Took some time over the week to play around with @swyx 's menubar source code and added in support for @perplexity_ai 's new Llama-2 Hosted runtime. It's so god damn fast, so curious about the source code behind it. Hopefully PR gets merged soon :)

2

4

20

@ivanleomk

Ivan Leo

4 months

Doing good rag is hard and full of tiny gotchas. Learn real actionable tips that have worked for our clients deploying RAG systems at scale. We're also capping max sign-ups with this course so that everyone gets enough air time to ask @jxnlco and other

Tweet media one

1

1

21

@ivanleomk

Ivan Leo

6 months

We got an event venue! If you've ever wanted to find out more about RWKV, here's your chance. We've got @picocreator from @RWKV_AI and @recursal_AI sharing about it and more over a round of drinks at @AWSCloudSEAsia Singapore. Shout out to @gabchuayz for helping to pull

1

6

20

@ivanleomk

Ivan Leo

26 days

@BowTiedFox What I found was useful was to 1. Take examples of copywriters you like 2. Get sonnet to extract writing guidelines rinse and repeat. Works decently well I would say.

1

3

21

@ivanleomk

Ivan Leo

2 months

@NeelNanda5 @3blue1brown @senr For a moment I thought 3blue1brown was Neel Nanda and my two worlds collided

1

0

21

@ivanleomk

Ivan Leo

9 months

There' s 3 ways to improve complex extractions using LLMS 1. use instructor w/ @pydantic 2. define relationships 3. incrementally update state now also available on NPM at @instructor -ai/instructor

1

3

19

@ivanleomk

Ivan Leo

4 months

I wrote up some of my thoughts after the AI Engineering Conference featuring some snippets from the talks i liked. Would love to get feedback on it! :)

Tweet card media

AI Engineering World Fair

Some thoughts from the AI Engineering World Fair

www.ivanleo.com

2

1

20

@ivanleomk

Ivan Leo

28 days

TIL that Gemini's cache has an unlimited TTL, stores up to 1M tokens and supports caching video, audio and even text. Seems like I'll really be switching over quite a few workloads to gemini @OfficialLoganK

Tweet media one

2

1

20

@ivanleomk

Ivan Leo

2 years

@fullStackRacc Lol why use typescript when you can have millions of users testing your code for free in parallel in production. Only cowards use type safety

0

1

20

@ivanleomk

Ivan Leo

4 months

SF has left me incredibly energized and happy with good discussions and the pleasant weather. I've been amazed at the quality of discussions here about all things LLMs and can't wait for the week to come :)

2

1

19

@ivanleomk

Ivan Leo

7 months

Built a lil cli with @tiangolo 's typer to ingest data and run some queries using @lancedb with the help of @jxnlco . Excited to ship more things out of this little cli

3

2

18

@ivanleomk

Ivan Leo

7 months

Interested in learning more about RWKV and what @recursal_AI is doing with this potential transformer killer? @aimuggle , @gabchuayz and I are organising a fun night of drinks with @picocreator - CEO of Recursal AI coming up in 2 weeks! Come by for good vibes and a great

3

4

19

@ivanleomk

Ivan Leo

1 month

The 1.5 billion free tokens with gemini have really been useful while I run my tests to verify code changes. Fun stuff coming to instructor soon!

3

1

19

@ivanleomk

Ivan Leo

11 months

if you're experimenting with open source models, you should definitely look into working with Modal's volume functionality. I spent some time working with it and I'm quite impressed. Works straight out of the box and provides a consistently fast experience for cold and warm

@ivanleomk

Ivan Leo

11 months

Spent some time in the evening playing around with @modal_labs 's new volume feature and it's definitely a huge step up. On @TheBlokeAI 's Open hermes model, I'm getting a speedup of around 10% in terms of raw inference speed for both Cold and Warm endpoints. Volumes also tend

Tweet media one

Tweet media two

2

2

25

2

1

18

@ivanleomk

Ivan Leo

9 months

I wrote down some reflections on my experience teaching myself ML. I'm still a beginner with much more to learn but this is basically what I've told most people who've asked me for advice. I wrote a longer article here:

@ivanleomk

Ivan Leo

9 months

I've had some people ask me for advice on ML so I compiled some of my thoughts into 3 tips that I wish I had thought of/done more of when I first started I'm honestly still a beginner but I figured I'd offer some thoughts as to how someone can juggle this with a fulltime job

9

34

280

1

1

17

@ivanleomk

Ivan Leo

6 months

Doing a presentation next week on a brief introduction to llms. I tried to condense some of the things I had learnt over the last ~ 6 months into a single article which could be read in < 15 mins at most

3

0

19

@ivanleomk

Ivan Leo

11 months

It’s a wrap on the meetup today organised with @eugeneyan , met a lot of new friends and caught up with old ones. Excited for what’s to come with another round of drinks and a paper club en route with @aimuggle and @SriniN123 in January

Tweet media one

5

3

16

@ivanleomk

Ivan Leo

1 year

Spent the day playing with @pydantic , Rich by @willmcgugan and selenium in order to build a small script to extract out all my saved google map restaurants. I'd say that ever since I started using Pydantic, input validation has never been easier

1

3

18

@ivanleomk

Ivan Leo

3 months

I'm thoroughly confused why there's vertex-ai and then there's the gemini api which has a completely different route.

9

0

19

@ivanleomk

Ivan Leo

4 months

The number of insanely interesting talks at the AI Engineering conference is incredible lol

2

0

18

@ivanleomk

Ivan Leo

8 months

Today I learnt that the C4 Dataset came from the Google T5 paper - that's pretty crazy ngl. This paper has so many gems honestly, if you're interested in learning more or discussing your thoughts on the paper, swing by the @latentspacepod discord later today for our weekly paper

Tweet media one

2

2

18

@ivanleomk

Ivan Leo

6 months

RNNs finally clicked thanks to @patloeber 's RNN tutorial, next steps is LSTMs but this tutorial series really helped me with pytorch. Everything makes sense lol

Tweet media one

2

1

17

@ivanleomk

Ivan Leo

11 months

Spent some time watching @dontusethiscode 's old pycon keynotes - So you want to be a Python Expert and it's phenomenal. I'd been avoiding generators and metaclasses for a while and I think this has convinced me to give them a second shot.

Tweet card media

James Powell: So you want to be a Python expert? | PyData Seattle 2017

www.pydata.orgPyData is an educational program of NumFOCUS, a 501(c)3 non-profit organization in the United States. PyData provides a forum for the internati...

www.youtube.com

2

3

18

@ivanleomk

Ivan Leo

8 months

Reading the BERT paper for the first time for the asia paper club run inside the @latentspacepod discord and it's a good reminder of how far things have come. Pretty crazy that back then they were blown away by 110-340M parameters and now we're looking at models that are

Tweet media one

2

2

18

@ivanleomk

Ivan Leo

9 months

3/ Read papers - I think this was very daunting for me and I've been doing it for some time. The benefit of most LLM research is that it's a lot more digestible than say traditional ML. People spend months of their lives to generate results and insights but you can get it in

1

1

17

@ivanleomk

Ivan Leo

1 month

With instructor 1.4.1, using Gemini is easy, even with Vertex. 1. Declare a Pydantic model 2. Use Gemini's multimodal messages 3. Enjoy automatic retries, validation and even streaming out of the box Thanks to @sonalsaldanha for the hard work implementing this!

3

2

17

@ivanleomk

Ivan Leo

2 months

You can get better results from your LLM by 1. Performing the task manually by hand first 2. Generating a few samples for few shots 3. Getting claude to rewrite your prompt after your first few rounds 4. Looking at the data Reminded of all 4 today after working with

1

0

17

@ivanleomk

Ivan Leo

6 months

We've put together a quick guide for you to get up and running with Logfire. Check it out now at

Tweet card media

Logfire - Instructor

Explore Logfire, an observability platform to enhance application performance tracking with Pydantic, Instructor, and OpenAI integration.

python.useinstructor.com

0

1

17

@ivanleomk

Ivan Leo

3 months

Generating large quantities of synthetic data at scale can get expensive. So we built out an easy integration for OpenAI's Batch Jobs using instructor

1

2

16

@ivanleomk

Ivan Leo

2 months

Man literally taught me how BERT and so many other models worked with his videos, so excited to see him on twitter

@hkproj

Umar Jamil

2 months

My new video on how to code a Multimodal (Vision) Language Model from scratch using only Python and PyTorch while explaining every single concept step by step. Link: #tutorial #pytorch #python #coding #fromscratch #llm #paligemma #gemma #visionmodel

21

155

925

0

0

15

@ivanleomk

Ivan Leo

4 months

First thing in the US is a big fat slice of pizza

Tweet media one

5

0

16

@ivanleomk

Ivan Leo

9 months

Fantastic turnout tonight at the AI Engineer Meetup! Thanks to everyone that came, was great to meet more people in the Singapore scene :)

Tweet media one

2

1

16

@ivanleomk

Ivan Leo

9 months

@cohere ’s new embed job api is very good actually, prob around 5-10x faster than using their previous API It took a bunch of sentences that I previously spent almost 45 minutes embedding and returned the final product in just under 5-10 minutes.

1

2

16

@ivanleomk

Ivan Leo

9 months

When you get rate limited by OpenAI and u crash ur entire batch job of 500,000 sentences even with 7x retries and backoffs of 3 mins

Tweet media one

3

0

16

@ivanleomk

Ivan Leo

8 months

Just finished another great paper club session covering BERT and here are 3 things I found super interesting about the paper. Thanks again to @swyx , @woof_hs and @bryanblackbee for dropping by and sharing about your experience working with BERT

Tweet media one

2

0

16

@ivanleomk

Ivan Leo

9 months

Use abstractions first, then throw them away. I'd also add here that it's ok to not know everything. I think this was something I struggled with a lot at the start - you've got a billion things to know and there's more to know/do the more you learn. Just compare yourself to who

2

1

16

@ivanleomk

Ivan Leo

3 months

Save more than 50% on your OpenAI bill today using just Pydantic and Instructor by running your first Batch Job. Use the same familiar syntax in Instructor that you've grown to love. This works for fine-tuned models too :)

Tweet card media

OpenAI Batch Jobs with Instructor - Instructor

A lightweight library for structured outputs with LLMs.

python.useinstructor.com

1

2

15

@ivanleomk

Ivan Leo

4 months

Thanks to @gabrielchua_ for having me! the talk can also be titled ‘Start with @lancedb full text search pls’ haha

@gabrielchua_

gabriel

4 months

@ivanleomk sharing on some advanced rag techniques maximising performance, moving beyond vibe checks, and how to make best use of @lancedb @the_builderclub

Tweet media one

Tweet media two

3

3

18

3

3

15

@ivanleomk

Ivan Leo

4 months

Probably read like 25 different papers just this week on prompting.. the range is so far. Some are just add this phrase while others build entire pipelines just to optimize the example choice. It's quite fascinating

2

0

15

@ivanleomk

Ivan Leo

7 months

Managed to get flash attention working in a jupyter notebook on @modal_labs . Thanks to @akshat_b for providing a useful code snippet.

Tweet media one

1

0

15

@ivanleomk

Ivan Leo

9 months

Working on implementing a simplified numpy version of Attention for the Paper Club this thursday, going to work on Multi Headed Attention Head tomorrow PR is open if anyone would like to give some feedback :

Tweet card media

Attention is all you need by ivanleomk · Pull Request #1 · ivanleomk/llm-paper-club-asia-notes

2

3

15

@ivanleomk

Ivan Leo

17 days

General thoughts on gemini after blowing 60 USD on it in the last 2 weeks 1. You need much more detailed instructions, especially for pro 2. Flash is a good model but when it comes to writing seems to output consistently shorter text

7

0

14

@ivanleomk

Ivan Leo

9 months

I'm a bit of a perfectionist so this took some getting used to. I think the biggest shift for me was to move from feature-complete to forcing myself to prioritise things that can be done in 1-2 weeks. Life often gets in the way and so making sure that you're working on smaller

1

1

14

@ivanleomk

Ivan Leo

5 months

@NielsHoven I think this is very true! I got into programming because I had incredible friends who showed me how fun it was. I imagine math might have been the same

1

0

13

@ivanleomk

Ivan Leo

6 months

Created a small pomodoro timer with NextJS backed up by D1 and deployed on @cloudflare with the help of @magicpatterns Thought I'd showcase some of the nice components they created

3

2

14

@ivanleomk

Ivan Leo

8 months

We r so back lol - pretty phenomenal lineup this week for the @latentspacepod discord's paper club We've got @wregss and @BhattGantavya presenting on their paper Matryoshka Representation Learning which Open AI uses under the hood for their new embedding model We've got

1

4

14

@ivanleomk

Ivan Leo

5 months

Ran my first 12k today, amazing start to the week

2

0

14

@ivanleomk

Ivan Leo

4 months

This was one of the best talks at AI Eng for sure

@ExaAILabs

Exa

4 months

How does Exa serve billion-scale vector search? We combine binary quantization, Matryoshka embeddings, SIMD, and IVF into a novel system that can beat alternatives like HNSW. @shreyas4_ gave a talk today at the @aiDotEngineer World's Fair explaining our approach! ⬇️

12

56

464

1

0

14

@ivanleomk

Ivan Leo

8 months

@jh3yy A bit surprised at some of the replies to be honest, it's not an uncommon question. it's 2024, figured we'd have moved past that lol.

3

0

14

@ivanleomk

Ivan Leo

30 days

Unlock the full potential of your audio and video content with Instructor's new list support for Gemini in 1.4.3 trying out the new pro and flash models has never been easier

Tweet media one

2

2

13

@ivanleomk

Ivan Leo

7 months

I've seen long bibliographies but 486 cited sources in a single paper takes the cake. Reading through this very comprehensive summary of LLMs over the past year with @bryanblackbee doing a recap tomorrow in the @latentspacepod . Can't wait to see how he's going to organise

Tweet media one

3

5

13