There are so many hackathons because there are so many people selling shovels

Eventually, codegen is the answer to everything.

'Maybe LLMs have been so successful because language itself has evolved over 1000s of years to be the scaffold in which human minds grow.'

Sholto Douglas: “People who work in other modalities have to put enormous amounts of thought into exactly what the right representation

Sometimes I just look at the wall and wonder what would it be like if RETRO worked like the authors claimed

arxiv.org/abs/2112.04426

I asked Claude 3 to generate an animation for the POV of a neuron in a large neural net. it gave me this:

Useful LLM application you could ship today:

Many articles by folks like Noah Smith 🐇🇺🇸🇺🇦 on substack etc. are filled with links for references

Ingest the linked pages and find the specific passages being referenced, then update the links to point to highlighted text

Google presents Genie

Generative Interactive Environments

introduce Genie, the first generative interactive environment trained in an unsupervised manner from unlabelled Internet videos. The model can be prompted to generate an endless variety of action-controllable virtual

ALOHA: learn fine-grained robotic manipulation via imitation learning

tonyzhaozh.github.io/aloha/

Learns policies for e.g. taking Airpods out of their cases etc. with under 10 minutes of human demonstration data

PCB design is an underrated application of code generation

One of the most tedious tasks that previously eluded automation

LLMs and traditional optimization techniques allow you to bring this to prod

The Python interpreter is an incredibly rich environment for self-play and synthetic data generation

OpenAI’s dominance in code generation is likely related to the fact that they are already heavily leveraging it internally

Many have beaten this dead horse but standardized testing is the *most* fair/equitable way of doing admissions

It’s shocking that higher ed (and magnet schools like Lowell in SF) were ever caught up in the psyop that removing it was somehow progressive

We’re super impressed by Andrew Gao Andrew Park molly cantillon ‘s hack at TreeHacks:

- sophisticated ML methods
- real/viable product vision
- creative use of latest tech

Our pleasure at Codegen to award them our “best use of agents” prize 🏆🏆🏆

Congrats!

The Codegen booth at TreeHacks goes hard

Come get a free chair massage while you hack

A common take has been”10mm context kills RAG”

Clearly wrong IMO

We have GPT-4 128k and yet it still makes sense to run RAG and retrieve a smaller number of samples in many scenarios

If we had a model that could span arbitrarily large datasets w/ inference cost similar to

Team Codegen is at TreeHacks for the next 48 hours!

Come say hey and learn about the bleeding edge of code generation at enterprise scale 🙌🤩

2/3 If you saw any discussion of math in California or the California math framework over the last few years, page 7 of the presentation is an absolutely amazing slide.

Here is the San Francisco district's own assessment of the changes they made: