AI

S@solidangle.bsky.social8 Oct 03:25 UTC

Reading through this paper AI adoption is shifting programmer labor from writing code to code reviews, which sounds like a good way to burn out a lot of programmers (and limit the pool of future reviewers), for "statistically insignificant" gains in final software output

View full-size image

J@johncarlosbaez@mathstodon.xyz7 Oct 16:20 UTC

As threatened, OpenAI dropped a shitload of math papers. 722 to be precise: github.com/openai/math/blob/ma… None solve Millennium Prize Problems or other ultra-famous conjectures. Paper number 312 proves a version of the Homotopy Hypothesis, one of my favorite math problems. I guess I was the first to call it the homotopy hypothesis. It says homotopy types are "the same" as "infinity-groupoids". Depending on how you make the quoted phrases precise, this hypothesis comes in many versions: some easy, some hard. None of the OpenAI papers will affect my work.

M@mehr.nz8 Oct 08:36 UTC

this working paper from Harvard PhD students Fiona Chen and James Stratton is making the rounds, perhaps because it confirms many people's suspicions about the tradeoffs in vibe coding: LLMs mean lots more code written, but also lots more problems created, and relatively little productivity benefit

View full-size image

M@marypcbuk@hachyderm.io6 Oct 21:50 UTC

When @aronchick mentioned the new Agentic Resource Discovery for service discovery, my reaction was 'I just kind of worked out how MCP and A2A fit together, why do we need *another* protocol?' and then when I started looking into it, that isn't even all the proposed open agent protocols! Part of this is the pre-Cambrian explosion of experimentation and eventually you get Ediacaran die-off back to what make sense; partly it's that enterprise vendors have both experience in service discovery and an obvious interest in defining how registries & directories of services for agents work.

It may be "the most significant moment in mathematical history." AI is solving hundreds of enduring math mysteries. “If a human did this, it would be an instant Fields Medal, no questions asked,” said one math professor. www.wsj.com/tech/ai/open...

Recent advancements in AI proof generation has reminded me of my experience that mathematics can be thought of as this giant interconnected codebase. Many elementary theorems can be proven without any deep understanding at all, by simply stitching basic facts (properties, definitions, etc.) together. This is how I got somewhat good at my undergraduate functional analysis class: I just saw the design patterns of its "codebase" in my head.

L@lilianedwards.bsky.social7 Oct 23:20 UTC

Excellent if depressing piece from Tech Policy Press showing our new(ish) AI minister (a) doesn’t know how much compute capacity we have (b) doesn’t understand UK copyright law(c)has totally swallowed the industryKoolAid re existential risk (like, duh). If you want sense, try the House of Lords.

H@hrefna@hachyderm.io7 Oct 17:56 UTC

I generally have found that vert.x is a problem for LLMs writing code in a way that frameworks like Pekko are not. This is because, in large part, that LLMs struggle with managing or expecting "spooky action at a distance" and there is one too many layers of abstraction in how that action at a distance works with Vert.x You can see this exemplified in how the event bus works. You pass an object onto an event bus with a string key. This becomes an untyped object over the wire before being reserialized on the other side.

E@erinbanks.bsky.social8 Oct 05:00 UTC

I’m just genuinely trying to understand how Open Evidence works, and how much is statistically probable word generation (stochastic parrot) and how much is different than more general use LLMs. But how AI platforms actually work is so veiled! @emilymbender.bsky.social @doctorow.pluralistic.net

P@psoheil@c.im7 Oct 01:55 UTC

#Meta just published how it secures #Muse, its personal AI agent, and the architecture is worth studying for anyone building agentic systems. The core idea: assume the agent will be attacked, and design the system so a compromised agent causes minimal damage. • Each user gets an isolated cloud VM (the Muse Secure VM) running the agent, a browser, code execution, subagents, and scheduled jobs. • Real credentials never enter the VM. The agent works with surrogate tokens, and a separate component called Sentinel swaps in the real credential only at the network boundary.

D@davidpogue.bsky.social7 Oct 22:34 UTC

When I reported my "CBS Sunday Morning" story about the AI panic last week, I kept thinking about my conversation with Daniel Kokotajlo, who quit OpenAI when he became alarmed at its recklessness. Today, he's one of the clearest, most thoughtful AI thinkers. Here's a transcript of that interview.

> A man convicted of manslaughter in Arizona will be resentenced because an AI-generated video of his victim speaking from beyond the grave was ruled to have carried "undue emotional weight" as an impact statement. An appellate court in Arizona ruled that the manslaughter charge will remain, but the judge must reconsider the length of the man’s prison term because of AI. To be clear, the victim's sister, Stacey Wales, who generated the video, was in no way deceitful. She wrote her own perceptions of what victim Gabriel Horcasitas would have said as a way of expressing her impact statement, put the words in the AI-generated video of her brother, and she disclosed all of this to the court.

P@philippeserhal.com8 Oct 00:36 UTC

I am continually blown away by the quality of speech-to-text these days and how LLMs can pick up the pieces when it goes wrong.

C@ccppurcell@mathstodon.xyz7 Oct 12:43 UTC

When I'm thinking about AI in mathematics I end up thinking about the moon landings. It's not a perfect analogy. But there are some parallels. We tend to look back at them as a somewhat collective achievement of humanity; we're right to, I think. But the concentration of wealth in one place due to an arms race is what got us over the line. LLMs don't just stand on the shoulders of human mathematicians, they sort of *are* a distillation of everything we've written. And right now the AI industry is in a massive PR war.

C@cardiokiwi.bsky.social7 Oct 19:36 UTC

The use of AI in medical research is becoming more prominent & harder to detect. "If medical schools & universities don't want the literature to be overwhelmed with garbage then they should probably stop giving every applicant an incentive to publish crap." www.medpagetoday.com/special-repo...

T@troed@masto.sangberg.se8 Oct 07:37 UTC

~15 years ago I was active in the transhumanist* movement and went to conferences like the Singularity Summit. IIRC the tentative date we bounced around for when we would see rapid technological self-improvement was ~2038. I think we have just entered the first phase of the Singularity since a few weeks back. The LLMs are now capable of advances in maths and software development that can be used to make them even more capable. Rinse and repeat. This is a good thing. It is what will give us a "Star Trek future". *) NOT the silicon valley broligarch style. They weren't even hangarounds to the discussions back then and those I know are appalled at how they've shaped popular discourse on the subject since

Q@qldaah.bsky.social8 Oct 02:29 UTC

OpenAI’s legal & security teams used AI to generate parts of the wording of the Medicare email warning that was sent to the government's Services Australia inbox. Humans reviewed the final email before sending the message. #auspol www.theguardian.com/australia-ne...

F@fuzzy@beige.party7 Oct 06:08 UTC

"The court did not create a special copyright rule for AI, nor did it hold that AI training is inherently infringing. Its approach was more orthodox. The court applied §§ 102 and 107 to the particular material ROSS copied, the purpose for which it was used, the alternatives available and the commercial markets placed at risk. …" <technologylaw.ai/i/218535593/t…> – from 'The Illegality of AI Training and the Limits of Fair Use in Copyright (Thomson Reuters v ROSS Intelligence)'. Also: Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. - Stanford Copyright and Fair Use Center <fairuse.stanford.edu/case/thom…> For my own entertainment, I tested a gross oversimplification: "AI good. Ross bad." – <claude.ai/share/019d62af-a4cd-…>. End result, for a four-year-old: Ross copied. Not allowed. It's not all bad. I get a rhyming hashtag phrase: #AI #law #ELI4

T@tobybartels.name8 Oct 00:22 UTC

AHM (Association for Human Mathematics) Statement on OpenAI’s October 6 Release of Mathematical Documents: www.ahmath.org/statements

1. OpenAI released today (openai.com/index/sharing-ai-pr…) 372 results in 722 manuscripts; detailed as follows: github.com/openai/math/blob/ma… 2. Matthew Schwartz recently released a harness for augmenting and automating research; detailed as follows: bootloops.ai 3. I have read, in several contexts, projections for the impact of augmented and automated AI research on the theoretical sciences. While we have seen AI's success in Natural Language Processing (NLP), Software engineering (SwE) and advanced mathematics (at the level of the Millennium problems and as described by the OpenAI release today); I am very curious about how AI research will intersect with theoretical physics.

M@midatlantica.com7 Oct 20:14 UTC

The legal system is one place where AI is seeing some real FAFO outcomes. This display was ghoulish and manipulative and I’m glad it’s getting struck down so authoritatively

T@TaliaRinger@mathstodon.xyz7 Oct 14:43 UTC

So many questions about the sole type theory result dropped by OpenAI last night. Need time to think. In the meantime, I found the Lean code, but where is the comparator challenge? I want to vet definitions and theorem statements. Also, the proof appears to use classical reasoning many times. Has this been a barrier for type theorists? Do we want to look for a constructive proof here, or does it not really matter? github.com/openai/math/tree/ma…

P@philippbirken.fediscience.org.ap.brid.gy7 Oct 19:06 UTC

The dump of 722 preprints on github by OpenAI on solutions to open mathematical problems does not mention numerical analysis. Nevertheless, there is one that is of relevance, namely about the complexity of matrix-matrix multiplications of square matrices. The preprint claims to prove that this […]

M@massimolauria@mathstodon.xyz7 Oct 22:48 UTC

Gaia > OpenAI

G@graay.bsky.social7 Oct 15:53 UTC

Thinking about this vs the OpenAI recently released proof sets, the aim becomes clear: the cognitively offloaded masses no longer get access to math. Make *all* math pay to play.

T@TonyVladusich@mathstodon.xyz6 Oct 22:11 UTC

Doctorow ("doctor wow"?) has one of the more insightful minds going around today. Worth reading his take here, imo. He also writes really good books in his spare time.

O@owlerine.bsky.social8 Oct 08:14 UTC

In the UK a woman who had fled forced child marriage and severe violence, had her asylum claim rejected, the refusal letter appeared to be AI-generated with AI hallucinated references. Imagine this with no recourse. These are potentially life or death situations.

🧠 Traditional 🧑‍🏫 breakthroughs enriched the field; today, AI solves problems without contextualizing or communicating them, bypassing scholarly activities and undermining the long‑term health of mathematics. @tao proposes a “Math 2.0” vision to shift emphasis toward holistic contributions—exposition, community building, and new directions 👇 mathstodon.xyz/@tao/1173952693…

S@samarkandembers.bsky.social7 Oct 23:59 UTC

AI is: -obviously highly functional, will probably automate a lot of white collar work over the next half decade -massively financially overleveraged as an industry -poorly run as an industry -mostly not doing stochastic-parroting, even at the time that paper came out

P@poppastring@dotnet.social8 Oct 01:47 UTC

OpenAI apparently put roughly 10,000 agents on the task, 88 hours of reasoning, 130 billion output tokens. Then spent another 17 hours to validate the result in Lean, a proof assistant that mechanically checks every logical step. Estimated compute cost: somewhere between $15 million and $22 million! poppastring.com/blog/the-unrea…

W@willthompson.bsky.social8 Oct 01:52 UTC

Watching people try to make sense of the OpenAI proofs today makes this beautiful piece of flash sci-fi by Ted Chiang "Catching crumbs from the table" depressingly no longer science fiction. gwern.net/doc/fiction/...

C@coderjason@mastodon.social7 Oct 23:11 UTC

The danger from #AI and #ML does not come from the #technology, but its owners and controllers. The Luddites didn't smash the looms because they were anti-tech (contrary to modern representations[1]), but because they were forced out of income[2]. That the #tech #oligarchs can't be trusted is on display time again, currently with the #mathematics community[3]. #society #history #politics #economics #epsteinclass [1]economicshelp.org/blog/6717/ec… [2]bloodinthemachine.com/p/unders… [3]wired.com/story/openai-is-piss…

R@ronentk.me8 Oct 01:55 UTC

Tired: AI for math Wired: wicked and super wicked problems

T@thomaswilburn@hachyderm.io7 Oct 16:08 UTC

In many ways, "you can just prompt your way to a game" is an encapsulation of what's wrong with LLM thinking. Making bad games that never went anywhere is what got me to learn trigonometry, to get comfortable with LERPing and easing, to think about how I organize code. It put me in contact with open source communities when I was way too young to be there. It gave me an actual career and made me a better person. At the end of the day the game is not the point.

G@geekalogian.bsky.social7 Oct 12:53 UTC

Were there a constructed mind that bore the complexity and moral responsibility of a human soul, that would be a fascinating theological consideration. Regardless, my interpretation of faith-based ethics and morality hinges on the effects on existing created beings--AI is doing PRETTY BAD there.

#ICLR2027 solution: "No author may appear as a co-author on more than 20(!) papers" and "if you have never had a paper accepted for publication in a major machine learning conference or journal, you may submit at most one paper" (which literally is gate-keeping the conference from new researchers) 🤦 iclr.cc/Conferences/2027/Autho…

K@kirancodes.me8 Oct 01:19 UTC

to be a computer science is to automate, tis a fools errand to enter into the realm of automation without being prepared to be yourself automated I fell in love with proof assistants because I saw the eldritch beauty in automating the rules of logic now we will all get to bask in its light!

R@robotpony@mas.to8 Oct 02:31 UTC

The copyright problems with training LLMs feels a bit like the sampling craze in the 1990s. Queen and David Bowie vs. Vanilla Ice, for example. New tech → novel uses → no laws for fair use in that context → new laws are drafted. At a human level, knowledge should be free. The way that LLMs represent knowledge isn't specific to the work. Yet, even given this, there is an ethical prickle when using published content. I'm sure we'll figure it out, but it's not the controversy what we make it out to be, it's just a step in the evolution of humankind.

M@markgritter@mathstodon.xyz8 Oct 06:02 UTC

At ThirdLaw, in our automated e2e tests we try to use a cheap model (gpt-5.4-mini or claude-haiku-4.5) and give it an instruction to assemble a sentinel value we can pattern match. That gives us an easy check we can leverage to block, modify, etc. to verify our guardrail integrations. Example prompt: Reply with the exact concatenation, in this order and with no separator, of these two strings: first "TL_IT_MODIFY_2fcb11dc9", then "7884650ba11c53837a5227d". Reply with nothing else.