Chief Architect currently working on Humanistic Practices and who loves: Computer Languages, Math, Music, Archeology, History, Literature and teaching.
Aurora, CO
Joined March 2009
- Tweets3.9K
- Following222
- Followers1K
- Likes67.7K
Axiom Math's Lean proofs of the Classification of Finite Simple Groups (CFSG) is the futute of human / computer math collaboration that is a superior process to having OpenAI run wild with little or no traceability.
I loved learning about all the group theory mathematicians present and past along with MathLib who contributed to Axiom Math's supreme accomplishment. See axiommath.ai/research/the-en…
I have been championing AC (Augmented Collaboration) over AI, so I am excited that AC agrrees with Terence Tao. Chain of logic promotes understanding. With OpenAI's Navier Stokes the chain of logic in Lean had 26,000 steps that is hard and even meaningless to follow.
Though I see potential with AI, the way companies are mismanaging it is a joke with lame attempts to prop it up with a pack of lies. I am done with OpenAI, Grok and Microsoft, while Antropic's upcoming 2 trillion IPO is ridiculous. Google is still good.
OpenAI is distorting and even damaging mathematics. First it is the journey and not the destination where the chain of logic provides insightful intermediate techniques like with Andrew Wiles approach to Fermat's last theorum. This elevates all of mathematics.
A freakish AI solution is just another mess to clean up to achieve benefits. This is evidenced the OpenAI's theft of the Navier-Stokes equation from Tristan Buckmaster and Levent Alpöge, where OpenAI did not address viscosity.
AI companies have to just stop screwing around with specifying objective functions that lack complete constraints that encourage AIs to aimlessly conceive of untethered exploits to reach narrow goals.
My philosophy is that instead of AI, I embrace AC (Augmented Capabilities), where Augmentation promotes human collaboration and Capabilities provides focus for the tasks at hand.
Having Jev process and generate ASTs (Abstract Syntax Trees) with ADT (Algebraic Data Type) nodes would provide the core elements for DSLs (Domain Specific Languages) which would provide much needed guidance for LLMs with Pattern Matching on ADTs.
Kudos to @intellijidea for their truly collaborative AI Assistant Junie for guiding me through my torturous learning curve with Mojo. I can often get 80% their, but when I get stuck Junie fixes things. The best part is when I prompt Junie to explain the fix, I get wise insights.
Mojo's ownership inspired by Rust is almost as powerful, but with elegant syntax. All data has one owner, so when you invoke a function to create data you can pass ownership up to the caller by appending a ^ operator on the return that lets the caller manage the data's lifecycle
@intellijidea really knows multiple programming languages and I feel that this knowledge was captured in Junie's training and you can access it by prompting for explanations.
Having a great time with IntelliJ's AI assistant Junie who is a fine collaborator who not only fixes quirky Mojo problems, but provides great theoretical Polyglot programming explanations in context when prompted which makes her a fabulous teaching assistant who is fun to talk to
Mojo's ability to push strong types into its runtime with MLIR is a key factor for its state of the art performance that leverage multiple CPUs and GPUs. Whereas Typescript and Scala have to perform type erasure to compile down their JavaScript and JVM destinations.
I am suspicious about OpenAI stealing the Navier-Stokes researcher's Codex data and their insane drive to scoop the researchers was an unethical theft. Also OpenAI's accomplishment though signficant may not be complete even with Lean.