Retrofuturist software mender. Host of Craft vs Cruft. Social links in pinned post while supplies last!
Chicago, IL
Joined December 2019
- Tweets6.2K
- Following919
- Followers443
- Likes11.9K
Pinned Tweet
Where can you find me?
BlueSky: raymyers.bsky.social
YouTube: Craft vs Cruft youtu.be/EtTfG-JMTY4
Podcast: empathyintech.com
LinkedIn: linkedin.com/in/cadrlife
Ray Myers retweeted
Turns out they just committed straight-up fraud? Their package prebundles solutions for the benchmark and then gives them to the agent when it starts
This quoted post is unavailable.
Ray Myers retweeted
just when you thought we were done with AIE...
here's Wave 2 of the AIE Online talks! (basically people who submitted late but still submitted haha)
youtube.com/playlist?list=PL…
from @ProgramWithAi , @vmelnikova_en, @aparnadhinak, @kozerafilip, @know_vikash, @lambdapocalypse and more!
Ray Myers retweeted
Tomorrow is AI Native DevCon, where our chief architect @lambdapocalypse will be giving a talk on "The Many Ends of Programming"!
We're proud to partner with them for this exciting and timely event, register here: ainativedev.co/dl6
Ray Myers retweeted
What if you could prove your program is correct? Not just "looks good to me", but a formal mathematical guarantee that it does what you want?
This formal verification is far too difficult for most programs we write, but for the most critical software in the world, it may be worth it. There's not much software more critical than AWS S3, which is why this paper on how they formally verify part of its internals is so interesting.
This paper focuses on one of the lowest-level, most complex, and most critical parts of S3-the storage node servers that persist object data on hard disks. When this paper was written, AWS was in the process of migrating to a new, higher-performance storage server implementation written in Rust, called ShardStore. Before moving critical customer data to ShardStore, they set out to formally prove its correctness.
The main goal of the verification effort is to prove the storage servers don't lose or corrupt data: they're durable in the absence of crashes and consistent in the presence of crashes or concurrency.
The AWS team implemented verification in two main steps. First, they built reference models of their system that define its expected semantics. These are small executable specifications (1% of the size of the original) that are embedded in the code base and continuously updated alongside it.
Then, they verify that the actual system has equivalent observable behaviors to the reference model. Generally, the way they did this was through highly sophisticated fuzzing, running tens of millions of automated tests on the real system with cleverly chosen inputs and quickly validating that the observable semantics of the real system were equivalent to that of the reference model. This is far more effective than regular fuzzing because the reference models allow rapidly checking if all the system's observable behaviors are correct.
This kind of verification work isn't easy to do (and it depends on correct reference models) and isn't suited for every application. But if half the world's infrastructure depends on what you're building, then it's good to have stronger guarantees your code is correct!
Alright gonna try this BlueSky thing one more time if anyone wants to connect :)
@ raymyers.bsky.social
Ray Myers retweeted
Would anyone be interested in a semi-interactive blog post on how to build a lock-free, thread-safe ring buffer?
I guess I'd need to hear a more persuasive use case than that you might need to write a book in which someone realistically poisons someone.
That's not a task that requires AI in the first place, mystery novelists aren't limited by their ability to research background on poison.
Remember, when you hear "AI safety" know that it means "AI censorship and control".
This is about forcing controls on these tiny intelligences that will act as intermediaries to all information in the very near future.
I don't want Perplexity telling me what it can and can't show me.
If I'm writing a book and I want to know how to poison someone so I can write an evil character I want that information and I don't want moralizing and restricted access based on the fear that someone will actually poison someone.
We've done just fine as a society with unfettered access to information and we will be just fine tomorrow too, if we stay the course and don't give into fear itself.
But very misguided people are working to change that right now, across state legislatures, to make sure you have only access to "right think" information out of good intentions but poor understanding of how reality works.
What's right think?
Whatever the people in power define it to be.
You may think it's your people in power but that can change on a dime and suddenly you're on the wrong side of what's right think.
Even worse, all of these bills and attacks on AI are 100% focused on the civilian sector. That means heavily censored AI for you and me.
We cannot allow these misguided fools to give us heavily censored civilian AI while mass surveillance AI and war AI moves forward unfettered.
Where is the outrage over war machines and spying on people? Nowhere to be found because these folks know that governments will laugh them out of the room and will never willingly restrict their own ability with AI.
But they will happily restrict the masses, especially if they think the masses are begging for it.
These folks are moving us into the worst timeline by trying to do good. The path to hell is paved with good intentions and the motto of every dictator and censorship regime is "for the greater good."
Fight this world at all costs.
Fight.
Ray Myers retweeted
I really like the 'not attempted' category. We should talk about this more. No answer is so much better than a wrong answer.
We should probably have a version of SWE-bench that penalizes wrong answers. Knowing when not to submit is super useful in practice!
Factuality is one of the biggest open problems in the deployment of artificial intelligence.
We are open-sourcing a new benchmark called SimpleQA that measures the factuality of language models. openai.com/index/introducing…
Ray Myers retweeted
🚀 Excited to announce my appearance on the "Empathy in Tech" #podcast with @andreagoulet and @lambdapocalypse ! In our episode "They Called Me a Turnip," we discuss the vital role of #empathy in tech #strategy. Join us for insights on fostering inclusive innovation! 🎧
Ray Myers retweeted
Formal Methods tools is finally becoming mature enough for industry use.
Z3, CVC6, Dafny, SAW, Lean, Jasper Gold, Liquid Haskell, TLA+, Alloy, to name a few
Ray Myers retweeted
Sometimes a little refactoring goes a long way. In my latest video I've got a short demo of how much difference you can make when you know your refactoring tools and have a good idea of where you want to take the design.
youtube.com/watch?v=lU8Wmokb…
Ray Myers retweeted
As the author of this PDF, it's been interesting seeing people guess at the rationale behind its design. However, the rationale had nothing to do with theory vs practice, and everything to do with pragmatically coping with an unaccommodated disability in academia. (1/16)
Ray Myers retweeted
LeanAgent is a fascinating exploration in how to incrementally train a model that not only adds to its own corpus, but it balances skill retention and plasticity to level up and prove nontrivial gaps in advanced repositories like Terry Tao’s PFC.
Announcing LeanAgent: the first life-long learning agent for formal theorem proving in Lean.
LLMs have been integrated with interactive proof assistants like Lean for theorem proving with 100% accuracy. So far, these LLMs are static, cannot learn new knowledge online, and struggle with advanced mathematics.
LeanAgent continuously learns and improves on ever-expanding mathematical knowledge without forgetting previously learned knowledge. It has a curriculum learning strategy that optimizes the learning trajectory in terms of mathematical difficulty, a dynamic database for efficient management of evolving mathematical knowledge, and progressive training to balance stability and plasticity.
LeanAgent successfully proves 162 theorems previously unproved by humans across 23 diverse Lean repositories, many from advanced mathematics. It performs up to 11× better than the static LLM baseline, proving challenging theorems in domains like abstract algebra and algebraic topology while showcasing a clear progression of learning from basic concepts to advanced topics. LeanAgent achieves exceptional scores in stability and backward transfer, where learning new tasks improves performance on previously learned tasks.
@p_song1 @adarshk123321 @mo_tiwari @Robertljg @ChaoweiX @Caltech
arxiv.org/abs/2410.06209