Wildroot · AI

Reading through this paper AI adoption is shifting programmer labor from writing code…

@solidangle.bsky.social — Thu, 08 Oct 2026 03:25:43 +0000

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

As threatened, OpenAI dropped a shitload of math papers.

@johncarlosbaez@mathstodon.xyz — Wed, 07 Oct 2026 16:20:57 +0000

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.

this working paper from Harvard PhD students Fiona Chen and James Stratton is making…

@mehr.nz — Thu, 08 Oct 2026 08:36:37 +0000

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

Recent advancements in AI proof generation has reminded me of my experience that…

@esamut@mathstodon.xyz — Wed, 07 Oct 2026 15:02:45 +0000

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. So what I am getting at here is that we should really stop pretending as if people can predict breakthroughs or estimate the importance/difficulty of problems. The project has been going on for some time, it has plenty of legacy code (not going to disclose what this corresponds to, I don't want to upset some folks) as well as some pretty elegant abstractions and templates (category theory, anyone?). Plenty of maintainers died along the way and their areas have fallen into obscurity. The whole thing is a mess! There was certainly some low-hanging fruit out there, we just didn't have the right tools to examine our codebase and see them. Well, until recently. Similarly to software engineers - mathematicians should evaluate each other based on the overall contribution, and not just per "ticket" (e.g. theorem). Refactoring math, as in simplifying things and making them more readable, is as important if not more important now than ever. This post will be updated, just wanted to share the rough concept first.

It may be "the most significant moment in mathematical history." AI is solving hundreds…

@demetriagallegos.bsky.social — Wed, 07 Oct 2026 18:43:59 +0000

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...

I generally have found that vert.x is a problem for LLMs writing code in a way that…

@hrefna@hachyderm.io — Wed, 07 Oct 2026 17:56:30 +0000

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. So in order to know what is happening the LLM has to know all of this and follow the request to the locations where the string key is read, and then figure out what it is doing with types on both sides. You can register types, but LLMs struggle to _remember to do this_. Then there is the matter of blocking. Setting up critical sections in vert.x requires an extra step that LLMs seem to forget constantly. In pekko you just have to enforce passing in a blocking dispatcher where this can happen and you are basically good to go. You can (almost) enforce this at compile time with a tool like archunit, and it is easy to catch otherwise, and it "just works." I do think you need to use a typed actor framework to really do this properly with LLMs, but that's not a _bad_ thing.

Excellent if depressing piece from Tech Policy Press showing our new(ish) AI minister…

@lilianedwards.bsky.social — Wed, 07 Oct 2026 23:20:49 +0000

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.

#Meta just published how it secures #Muse, its personal AI agent, and the architecture…

@psoheil@c.im — Wed, 07 Oct 2026 01:55:15 +0000

#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. • Sentinel is the only thing that can talk to connectors or the internet. The agent proposes, Sentinel allows, denies, or asks the user. The agent cannot override it. • Tainted egress: once a process touches private data, it loses automatic network permission. Any later outbound action needs human approval. This is Meta's answer to the "lethal trifecta" of private data, untrusted input, and a way to send data out. • The model is trained with prompt injection in mind, but Meta's position is that model-level defenses are never enough, so the real controls live at the OS level. Notably, Meta also opened a public bug bounty for Muse: up to $300,000 per report, including up to $130,000 for a successful prompt injection affecting a single user. The broader lesson: permission needs an owner outside the agent that wants to act. #AIAgents #AISafety #Cybersecurity #PromptInjection #Meta research.meta.ai/blog/security…

I’m just genuinely trying to understand how Open Evidence works, and how much is…

@erinbanks.bsky.social — Thu, 08 Oct 2026 05:00:20 +0000

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

> A man convicted of manslaughter in Arizona will be resentenced because an…

@mitch@hachyderm.io — Wed, 07 Oct 2026 01:11:00 +0000

> 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. > She compared it to the way courts thought of photography in the late 19th century. > "It took about 15 years […] and about five landmark cases in the United States that went all the way up to the Supreme Court before photography was an accepted standard to be used in the courtroom for evidence, identification, and testimony, et cetera,” she said. “I believe that's what we're seeing now with AI. It is a brand new medium.” 404media.co/her-ai-generated-v…

When I reported my "CBS Sunday Morning" story about the AI panic last week, I kept…

@davidpogue.bsky.social — Wed, 07 Oct 2026 22:34:29 +0000

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.

When I'm thinking about AI in mathematics I end up thinking about the moon landings.

@ccppurcell@mathstodon.xyz — Wed, 07 Oct 2026 12:43:16 +0000

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. I think it's at least plausible that this period ends rather quickly, as the moon landings did. I wonder how many articles were written in 1969 about how we'd all be travelling to the moon one day. Another parallel is: what is the inherent value of putting someone on the moon? I think it's quite similar to a landmark theorem. It broadens our horizons, maybe even enables further research, but it doesn't have much to do with the price of fish. #ai #math #llm #science

I am continually blown away by the quality of speech-to-text these days and how LLMs…

@philippeserhal.com — Thu, 08 Oct 2026 00:36:08 +0000

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.

~15 years ago I was active in the transhumanist* movement and went to conferences like…

@troed@masto.sangberg.se — Thu, 08 Oct 2026 07:37:43 +0000

~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

in a rational world, the second paragraph here would be about how Dave was dismissed,…

@ptfen.bsky.social — Thu, 08 Oct 2026 09:19:47 +0000

in a rational world, the second paragraph here would be about how Dave was dismissed, and development was focussed on expanding the platform's robustness and capacity for 4K content, but here we are

"The court did not create a special copyright rule for AI, nor did it hold that AI…

@fuzzy@beige.party — Wed, 07 Oct 2026 06:08:33 +0000

"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. …" – 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 For my own entertainment, I tested a gross oversimplification: "AI good. Ross bad." – . 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

OpenAI’s legal & security teams used AI to generate parts of the wording of the…

@qldaah.bsky.social — Thu, 08 Oct 2026 02:29:21 +0000

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...

1.

@ScottCaveny@mathstodon.xyz — Wed, 07 Oct 2026 01:34:44 +0000

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. Aspects of quantum mechanics (field theories, computing, and foundations) pose challenges that appear, to me, potentially unique to the domain of theoretical physics where the successes of formalization (through chains of reasoning or Lean verification) seen in NLP, SwE and advanced mathematics may not apply as directly or as immediately. I am certain that we are about to learn what happens when LLM's collide with theoretical physics

AHM (Association for Human Mathematics) Statement on OpenAI’s October 6 Release of…

@tobybartels.name — Thu, 08 Oct 2026 00:22:57 +0000

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

So many questions about the sole type theory result dropped by OpenAI last night.

@TaliaRinger@mathstodon.xyz — Wed, 07 Oct 2026 14:43:10 +0000

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…

The legal system is one place where AI is seeing some real FAFO outcomes.

@midatlantica.com — Wed, 07 Oct 2026 20:14:08 +0000

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

Gaia > OpenAI

@massimolauria@mathstodon.xyz — Wed, 07 Oct 2026 22:48:42 +0000

Gaia > OpenAI

The dump of 722 preprints on github by OpenAI on solutions to open mathematical…

@philippbirken.fediscience.org.ap.brid.gy — Wed, 07 Oct 2026 19:06:34 +0000

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 […]

🧠 Traditional 🧑‍🏫 breakthroughs enriched the field; today, AI solves problems without…

@loleg@hachyderm.io — Thu, 08 Oct 2026 06:11:57 +0000

🧠 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…

Thinking about this vs the OpenAI recently released proof sets, the aim becomes clear:…

@graay.bsky.social — Wed, 07 Oct 2026 15:53:26 +0000

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.

OpenAI apparently put roughly 10,000 agents on the task, 88 hours of reasoning, 130…

@poppastring@dotnet.social — Thu, 08 Oct 2026 01:47:56 +0000

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…

In the UK a woman who had fled forced child marriage and severe violence, had her…

@owlerine.bsky.social — Thu, 08 Oct 2026 08:14:51 +0000

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.

The danger from #AI and #ML does not come from the #technology, but its owners and…

@coderjason@mastodon.social — Wed, 07 Oct 2026 23:11:43 +0000

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…

AI is: -obviously highly functional, will probably automate a lot of white collar work…

@samarkandembers.bsky.social — Wed, 07 Oct 2026 23:59:35 +0000

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

In many ways, "you can just prompt your way to a game" is an encapsulation of what's…

@thomaswilburn@hachyderm.io — Wed, 07 Oct 2026 16:08:43 +0000

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.

Watching people try to make sense of the OpenAI proofs today makes this beautiful piece…

@willthompson.bsky.social — Thu, 08 Oct 2026 01:52:33 +0000

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/...

#ICLR2027 solution: "No author may appear as a co-author on more than 20(!) papers" and…

@djoerd@idf.social — Thu, 08 Oct 2026 08:18:20 +0000

#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…

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

@ronentk.me — Thu, 08 Oct 2026 01:55:41 +0000

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

The copyright problems with training LLMs feels a bit like the sampling craze in the…

@robotpony@mas.to — Thu, 08 Oct 2026 02:31:42 +0000

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.

Were there a constructed mind that bore the complexity and moral responsibility of a…

@geekalogian.bsky.social — Wed, 07 Oct 2026 12:53:32 +0000

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.

At ThirdLaw, in our automated e2e tests we try to use a cheap model (gpt-5.4-mini or…

@markgritter@mathstodon.xyz — Thu, 08 Oct 2026 06:02:03 +0000

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. (For non-guardrail tests we can just synthesize both request and response-- these have to go through a real model router. In some cases we could mock the provider's API on the other side, but don't do so at the moment.) But, we've discovered that these models pretty consistently fail at this task, enough that even with retries we can't get a clean test run. Common symptoms are doubled or omitted characters and weirdo separators (everything from - and _ to space to one of the Unicode zero-width spaces). Sometimes it will put in the undesired separator someplace else in the string than the original break point! I ran an experiment tonight testing the hypothesis that this task might be easier using real words and word boundaries rather than hex. The results are 76/80 success rate with the hex string, 75/80 using an arbitrary break point using words, and 80/80 using word boundaries. On a day where OpenAI dumped a huge trove of sophisticated math output it's kind of strange to be dealing with "can you even concatenate strings as asked" :)

to be a computer science is to automate, tis a fools errand to enter into the realm of…

@kirancodes.me — Thu, 08 Oct 2026 01:19:34 +0000

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!

It’s funny to hear developers whine about AI-related job loss.

@Scottw@ruby.social — Wed, 07 Oct 2026 14:21:52 +0000

It’s funny to hear developers whine about AI-related job loss. Software typically exists to do things humans would have done manually (and been paid to do). Anyone who’s been a professional engineer long enough has effectively eliminated thousands of jobs. I don’t see a future with fewer engineers than today, but the barriers to entry and the optimal skill sets are changing. That will push some folks out, but it’ll open the door to many who just never groked code.

With unchecked AI deployment posing severe risks to democracy, security, and human…

@Fife4Europe@mastodon.scot — Wed, 07 Oct 2026 13:03:04 +0000

With unchecked AI deployment posing severe risks to democracy, security, and human rights, statutory guardrails are vital to protecting public interest. While the EU sets enforceable safety standards, voluntary corporate promises leave society exposed. euronews.com/2026/10/07/laws-p…