Unhinged Education for an Unhinged World
Multiverse
Canvas mode melts the curriculum.
Liz reviewed OpenAI Dev Day, started updating the Using Language Models curriculum for Canvas, and noted the first Prompt Engineering run was nearly done.
See upcoming classesThis issue is dressed as Op Art
OpenAI Dev Day happened, and Liz shared the keynote along with her take on it. She started updating the Using Language Models curriculum to cover the new Canvas mode.
The first run of Prompt Engineering was nearly finished, with Swarm Architecture as the remaining class. The track page lists the full sequence.
Liz also said she would hire someone to answer student questions, and that a break week was coming.
Hey Multiverse Crew-
Whew, lots of changes this week. Elena visited, Roan went back to the UK, and I got to meet back up with a bunch of friends I lost contact with over the years. Thrilled to see folks. I did a Corporate Gig on Oct 1-3, which helps me to pay the rent.
OpenAI Dev Day happened, and almost everyone from the original OpenAI has quit, due to Sam Altman being Littlefinger (I hate it when Musk says a true thing but stopped watches are right twice a day etc). I'm working on an update to the Using Language Models curriculum to cover the new Canvas mode.
If you've been waiting for my Autogen Class its called Swarm Architecture - it got pushed because of low enrollment. Sign up here. That means we're almost done with the first run of all of the Prompt Engineering classes! Haha, it took 9 months and most people didn't make it all the way through because they were overwhelmed and we needed to pretty much double the length of the classes by the 3rd cohort because senior developers and data scientists are struggling to keep up.
If you're waiting on an email from me... just wait longer sorry 🏊🏊🏊 I will hire someone to answer questions soon, it's ridiculous right now. I am so behind on everything, I'm excited to get a break week next week! So, in honor of taking time off, I'm headed to Detroit to take absolutely no time off. I'll sleep when I'm dead! Dead I say! 😴🪦
Revised to cover Canvas mode.
Are you interested in the concept of Intelligence research, AGI, Psychometrics and are absolutely positive your unique field and set of interests will allow you to meaningfully contribute, but are not an I/O Psychologist or psychometrician?
You want to attend prompt engineering! You'll learn all about the process of building effective benchmarks there. We'll be discussing the research, and you'll learn quite a bit more in that class. You can help by attending this class and attempting to make your own benchmarks, this is literally the research. We'll discuss the forms of validity in that class, and you are able to contribute to the research by constructing your own benchmarks and having them reviewed. This will help us.
23 And Me Systemically Erases Indigenous Families So I Am Glad The Company Is Dying Out, I Hope Someone Tells Whatever Her Name Is She Doesn't Exist As A Response Because She Doesn't, She Was Never A Real Person In The First Place, She's Just A Myth Some Families Tell
Tipping Points in the Climate are going in Both Directions
Recording Jammers are mostly BS
Everyone Is Doxxed
China's Robots Look Cooler Than Our Robots
Uhh, The Matrix I Guess
Here's What Happens Next
StoryMaker, a framework to keep consistent characters across generated images. With consistent clothing and facial reproduction. Uses learnable background embeddings. Lots of examples of use cases. IP adapters, LoRAs, ControlNets etc feature heavily.https://t.co/iBcpdkEbHu
— cycloarkane (@cycloarkane) September 27, 2024
Cool novel RAG approach, using lightweight LLMs as memory modules with massive contexts while the heavy cognitive lifting is done by another model. The authors made their own UltraDomain benchmark which focuses on long contexts to test their system.https://t.co/ywLhHbEBC9
— cycloarkane (@cycloarkane) September 25, 2024
Kolmogorov-Arnold Transformer (KAT), a combination of KANs and transformers. Superior performance on on ImageNet1k compared to traditional MLP transformers, more scalable. KAT is more scalable than simple KANs as it uses group KAN shared functions.https://t.co/OR2XlmU2LH
— cycloarkane (@cycloarkane) September 21, 2024
V-STaR. Using incorrect reasoning and correct reasoning to gain even more substantial improvements through verification. Using a correctness label allows learning from bad reasoning. Authors only used LoRAs so full SFT may yield even better results.https://t.co/KpHcvksPtb
— cycloarkane (@cycloarkane) September 17, 2024
Efficient LiDAR odometry that reuses data and uses adaptive local map updates to limit the drift. I was wanting to explore LiDAR a bit more so I can understand Waymos better tbh lmao. C++ code available with the paper.https://t.co/kcsLSOWHAz
— cycloarkane (@cycloarkane) September 16, 2024
Paper on censorship and inferential adversaries. Bit draconian imo but good to know what might be around the corner. Basically about censoring information that could be aggregated to do bad stuff but isn't in itself disallowed.https://t.co/Tmjwc92qmr
— cycloarkane (@cycloarkane) September 15, 2024
More sampling can increase success by 300% and out perform SOTA models. This paper is a must read to understand 🍓. Implementations of this could be insane for OS models power on local compute, especially if you have time to do the computations.https://t.co/HxwtMsuxdJ
— cycloarkane (@cycloarkane) September 14, 2024
If you don't have time to read the OpenAI GPT-o1 (🍓) paper I read it myself and put key insights in a markdown bulletpoint list. Some really wild stuff in there and a pretty candid analysis from OAI it seems.https://t.co/aJYLmlI1gk
— cycloarkane (@cycloarkane) September 12, 2024
Pick a learning path. Recordings and curriculum, or live class seats too. Every class is pay-what-you-can, and scholarships are open to anyone who needs one.
A new issue every Sunday. Follow the RSS feed, or come make something for the next one — see what’s running.
