In honor of the playoffs, I’d like to showcase what we’ve been working on here at Nexavision — a new way to generate basketball analytics through tracking with computer vision and AI: 🧵
Proud to announce winning the Sports Tech hackathon at
@fdotinc
! Came up with a novel way to track players on a NBA court with any video angle, excited to see its applications!
We’ve built a system that can take in any basketball video feed to then track and identify all the players on the court. Most importantly, we calculate probabilities of making shots based on where the players are at, their movements, who the players are, and all the defenders.
Curious to check out all this? Come visit us at to check out our stats! We are covering the rest of the NBA playoffs, WNBA coming soon, and we’ll get the NCAA back in the winter.
We’ve taken this data and put it up on our site at NexaOdds so you can view all the shots taken after a game, their probabilities, and a whole bunch of other data on different possessions.
Are you a full stack dev interested in building player tracking software for basketball? Want to work with different collegiate+professional teams and help us build the front end for a new way of generating analytics? Send me a DM, we are looking for a cracked builder.
We're excited to officially welcome Nexavision to the family 🚀
They're changing the way we track and understand basketball.
Drop a 👋 to welcome
@AmarSVS
!
It was an honor to present Nexavision at the MIT Sloan Sports Analytics Conference this past weekend. We really wanted to showcase the amount of data you can generate with only broadcast video, and everyone loved to see the tracking demos. Can’t wait to see where things go from
@chrisalbon
colab/kaggle for light iteration, lambdalabs is the cheapest gpu provider, runpod is in and upcoming, gcp/azure/aws if you just want to spin up a gpu powered vps and then ssh tunnel to a jupyter notebook
I was holding my two year old daughter on a walk moments ago and for the first time out of nowhere she said,
“GPT4 is just a stochastic parrot.”
I started to quietly tear up. She couldn’t see my face and I didn’t make a sound, but she said,
“It doesn’t actually reason”
I was holding my two year old daughter on a walk moments ago and for the first time out of nowhere she said,
“God is good. God is real.”
I started to quietly tear up. She couldn’t see my face and I didn’t make a sound, but she said,
“Don’t cry. It’s okay.”
@analyticsaurabh
@HamelHusain
8bit/fp16 causes issues with flan t5 models, probably better to go with one of the bajillion decoder only models being pushed out rn
Data augmentation through similarity searches from model embedding has been out for a while now. This only offers a small increase in performance because the augmented data is not diverse. I’m sure you can do better on benchmarks by just asking the model to critique itself.
Simple Self-Improvement of Code LLMs
1) Pre-train & Fine-tune code LLM, gaining knowledge
2) LLM then generates pseudo outputs
3) Add that to original data & train for next epoch
Significantly improves code summarization & code generation performance
What happened to MegaBYTE? Did anyone ever scale it? Are we just at the point where we get excited about new architectures every week and forget them the next because nobody actually investigates further?
If you ever are building a lengthy data analysis project and need fast iteration, consider building a modular pipeline. The little extra time of building a generalizable platform can dividends when testing multiple approaches.
@RoxCodes
Generate embeddings for a body of text, isolate sentences or paragraphs that share high similarity with the whole text but not with each other, use an LLM to summarize that. Cheap, effective, and fast (combo of old school NLP + modern techniques)
@KennethCassel
CLIP, Flamingo, with new research coming out about other joint LLM-ViT models. Check out the kaggle stable diffusion to prompt competition for cutting edge open source image to text models being built
Amar was our first place winner of ArenaX — an hackathon
From med school dropout to working with NBA teams in < 1 year.
ArenaX was his ticket to the future, maybe DreamXR can be yours.
Sign ups close this week:
Sounds cool until you realize that the tokens from twitter aren’t necessarily providing a ground breaking advantage… google has much more valuable data and already a better ML support team ready and look how launching AI went for them
💡 In summary, Elon Musk's acquisition of Twitter, the birth of X AI, and the changes to Twitter's algorithm all point to his plan to revolutionize AI.
And YOU, the users, are at the heart of it all.
But are you comfortable with your tweets fueling Elon Musk's AI project? 🤔🐦
Try taking a coffee nap. Drink a cup of coffee, or any source of caffeine, and immediately fall asleep. This may sound funny, but it's a proven way to increase focus and wakefulness. How does it work exactly? 2/n
While this sounds like fantasy and there are things I do not agree with (surgical robots are still miles away and a completely different task, and text to protein will likely be done with autoregressive generative models instead of diffusion models), this is not a far future.
The next phase of AI for medicine will be about *Generalist Medical AI* models.
Versatile in ability, multimodal in design, and conversational in behavior, GMAI will be a major shift in how we develop and deploy.
Our new Nature paper outlines the vision:
Lots of talk about Devin… won’t be replacing engineers anytime soon but stacks like this are going to make developing a whole lot easier. The LLM development stack is slowly maturing and the next logical step is including multi source inputs (terminal, IDE, browser) as part of
i just quit my job!!
i'm in love with storytelling.
so now i'm officially launching my visual podcast series - harpriya talks.
i'm recreating the vibe of having a deep convo with a friend on a bedroom floor at 3am.
hot takes and intimate thoughts.
Colab can sometimes offer better access to data if you upload it to your drive, and for only $10/mo you can use A100s for training, it is still a solid option imo
2. Presence and Frequency Penalty:
These both affect tokens that are either frequent in the prompt or in the training data. Increase presence_penalty for creative outputs, and increase frequency_penalty for more concise outputs
Integer division is exceptionally slow on a GPU, so I've written a single-instruction, O(1) complexity integer division algorithm in CUDA. Please feel free to use it (with attribution).
1. Temperature and top_p:
These are both measures of "randomness" or "creativity". If you want more creative text, increase the temperature. For more accurate generations, lower the top_p.
Introducing Bifrost, a tool that uses AI to turn anything from Figma into clean React code automatically.
How it works:
- Generates usable code for anything in Figma
- Drops the code directly into your code editor
- Learns how to structure code just like yours
1/4
True - I printed my ML model vectors onto a piece of paper and deleted them from my hard drive so no one can hack them from me. I’m now self hosting a mail powered API - mail me your inference request and I do the matrix multiplication by hand and mail back the result
3. Logit Bias:
Directly add a bias onto certain tokens that you would like to be more or less prevalent in your responses. Penalize toxic outputs and if you want a company name to be repeated -- you can influence the generations with a positive bias to do that
@MelindaBChu1
@TrevorCampbell_
@tlbtlbtlb
I believe they were talking about software viruses/malware, which are valid risks considering its much easier to test them out. It’s difficult enough to sort through TBs of data, but even then since they are mostly engineering principles, I believe RL is the way to go
@bradneuberg
This is the essential problem — you could likely do tricks like use knowledge graphs and only train LLMs to produce summaries of them to limit them to reasoning. Otherwise the weights of LLMs intertwine reasoning and knowledge so they can’t be separated
@sathaxe
@AmarSVS
’s sleep stack:
- Pound back a shot of espresso
- Lay head back in office chair and nap for 15 minutes before the caffeine kicks in
- Wake up and keep hustling out AI models for Bifrost
Adenosine is released when you take a nap, so if you immediately lie down after drinking some coffee and sleep for ~20 minutes, you enable the caffeine stimulant to provide its optimal effect to keep you up 4/n
There is a molecule called adenosine that accumulates as you stay awake. It contributes to making you feel tired, and caffeine works by serving as a competitive inhibitor to the receptor that binds adenosine 3/n
@bradneuberg
Current meta is to use retrieval to provide appropriate context and only use LLMs to reason based on that — however context size and appropriate retrieval methods are bottlenecks
@bradneuberg
Attempts have been made such as neural Turing machines and “LLMs are computationally universal” - use LLMs only for reasoning and store results externally. These approaches suffer hard limitations however, it is difficult to join external knowledge efficiently with attention
If you aren’t using an on prem GPU workstation, what’s a better solution than spinning up/down cloud VPS’s and jupyter notebooks and developing code through that? Surely some startup has made a better workflow?
Transformers will give the breakthrough that will allow us to automate some white collar jobs. Reinforcement learning will give the breakthrough that will allow us to create intelligence smarter than humans.