AI proves an Erdős conjecture & Data filtering in AI pretraining - AI News (May 22, 2026)

Let’s start with that math headline. OpenAI reports that an internal, general-purpose reasoning model produced a proof that overturns a long-standing Erdős conjecture tied to the planar unit distance problem—basically, how many pairs of points can be exactly one unit apart as the number of points grows. For decades, the intuition was that familiar grid-like constructions were close to optimal. This new result claims a stronger construction exists, creating infinitely many point sets with meaningfully more unit-distance pairs than expected. What makes this especially notable: external mathematicians reportedly checked the AI-generated proof and published a companion explanation, and the argument pulls in heavy machinery from algebraic number theory—tools you wouldn’t expect to show up in what looks like a simple geometry question. For AI, this is a rare public example of a system not just assisting, but generating a verifiable path through an open research problem.

On the more “how we build models” side, a new arXiv paper by Mohri, Duchi, and Hashimoto challenges a popular assumption in pretraining: that you must aggressively filter data for quality. Their scaling experiments suggest that in a compute-rich, data-scarce regime, the best filter can be no filter at all. In other words, larger models trained longer may tolerate, or even benefit from, the messy long tail of data many pipelines would normally discard. If this generalizes, it could change the economics of model building—less obsession over perfect curation, more emphasis on whether you can afford the compute to let the model sort signal from noise.

Staying with efficiency, researchers from Google DeepMind and Seoul National University introduced LiteFrame, a video encoder meant to make video-focused LLM systems handle much longer clips without exploding inference cost. The key claim is that once you’ve already reduced tokens going into the language model, the vision encoder becomes the real bottleneck—because it still has to grind through frames. LiteFrame’s promise is a better latency–accuracy trade-off, letting systems process far more frames under the same budget. That matters because long-video understanding is where a lot of “agentic” usefulness lives—meetings, lectures, gameplay, security footage—yet it’s been brutally expensive to do well.

We also saw another push toward unified multimodal systems with ByteDance’s open-source Lance, a 3B-parameter model aiming to understand, generate, and edit across images and video in one framework. The headline here isn’t that it beats every giant model—it’s that teams keep trying to collapse fragmented toolchains into a smaller, single model that can both “see” and “create.” If that trend holds, it could lower the barrier for developers: fewer specialized models to glue together, and fewer separate failure modes when a product needs both understanding and generation in the same workflow.

Audio generation had a busy cycle too. Stability AI released Stable Audio 3.0, emphasizing open weights and—importantly—licensed training data and clearer commercialization terms. Whether you buy the licensing framing or not, it’s an explicit response to the legal uncertainty that has hung over generative music. In parallel, Meta’s Facebook Research published WavFlow, a research release that generates audio directly in raw waveform space from video and/or text, instead of leaning on compressed latent audio codecs. They didn’t ship their production checkpoints, but the research point is provocative: high-quality synchronized audio may not require the usual compression tricks, which could simplify future pipelines.

Now, the money and compute story that keeps getting bigger: a securities filing says Anthropic has agreed to pay SpaceX nearly 45 billion dollars over three years for compute to support Claude—roughly framed as billion-plus monthly payments, with some ramp timing and termination flexibility. Two takeaways. First, for leading AI labs, compute isn’t just an expense line—it’s the constraint that shapes product strategy. Second, SpaceX is positioning itself as an AI infrastructure supplier, not merely a launch and satellite company. If investors were looking for durable, non-space revenue streams ahead of an IPO era, this is the kind of disclosure that changes the conversation.

At the same time, enterprises are getting louder about inference costs. A report argues that companies like Meta, Shopify, Spotify, and Pinterest are feeling LLM expenses as a margin drag. And it raises a pointed question: can the top AI vendors sustain premium pricing when strong alternatives are proliferating? The picture being painted is a “hybrid” enterprise stack: cheap or open models handle the bulk of work, and frontier models get called only when they’re truly needed. That’s a rational engineering decision—but it’s also a direct threat to the revenue assumptions that blockbuster IPO valuations rely on.

Related to that, another analysis compares API pricing moves across Google, OpenAI, and Anthropic and reads them as a shift from growth mode to profitability mode. Prices have moved around enough to suggest that at least some earlier pricing may have been strategic—subsidize adoption, then tighten the screws when usage is sticky. The broader implication is simple: as AI capex keeps climbing, “cheap AI forever” was never guaranteed. If you’re building on these APIs, cost volatility is now part of the risk model, not an edge case.

Speaking of IPOs, the Wall Street Journal reports OpenAI is pushing forward with plans to go public, potentially as soon as September, after Elon Musk lost a lawsuit that had threatened the company’s structure and finances. Banks are reportedly lined up, and a confidential filing could come soon. If it happens, it won’t just be a big tech listing—it’ll be a referendum on whether public markets believe AI’s revenue can eventually catch up to its infrastructure spending.

On the hardware chessboard, Alibaba unveiled an in-house accelerator called the Zhenwu M890, framed as part of reducing dependence on Nvidia amid tightening U.S. export restrictions. The chip is pitched around “agent” workloads—tasks that involve long context and lots of coordination—which conveniently aligns with where big customers are trying to go next. Zooming out, this is another signal that geopolitics is hardening the AI supply chain: more domestic silicon roadmaps in China, more long-term compute contracts elsewhere, and a general move away from assuming the global GPU market will stay frictionless.

Now for the human and incentive side of AI—because the backlash stories are getting sharper. One blog post by Axel argues that today’s generative AI ecosystem amounts to large-scale, unauthorized plagiarism: models trained on writers’ work without consent, and downstream users prompting tools to reproduce tutorials and then publishing the results for profit. Axel points to a personal example where AI-generated pages allegedly mirrored his e-commerce tutorial content closely enough to preserve links back to his site—yet those copycat pages ranked higher on Google. Whether every detail holds up or not, the bigger issue is real: attribution, training data rights, and search incentives can combine to reward mass replication over original work.

That tension shows up inside engineering teams too. A post from Evil Martians describes “AI burnout” as AI-assisted workflows compress work into intense cycles of prompting, reviewing, and debugging—often leading teams to do more, not less. The claim is that craft satisfaction drops, review bottlenecks grow, and people lose time to think slowly, which is exactly when you catch subtle risks. The message isn’t anti-AI so much as anti-expectation: if management treats AI as a permanent speed multiplier, the result may be faster exhaustion and higher operational risk.

And culturally, the tone around AI is shifting in public settings. Steve Wozniak, speaking at a commencement, got cheers for joking that graduates have “AI—actual intelligence,” nodding to anxieties about automation. Meanwhile, other speakers have been booed for more forceful “embrace AI or else” messaging. In a separate media-business angle, BuzzFeed’s Jonah Peretti is stepping down as CEO and taking a role branded around AI as the company’s ownership shifts. Critics frame this as another case of “AI will save us” repositioning without clear evidence. Taken together, it’s a reminder that AI adoption isn’t just technical—it’s tied to trust, jobs, and whether institutions look careful or careless.

Finally, a small but telling web signal: Google added an experimental check for an llms.txt file in Chrome Lighthouse’s new “Agentic Browsing” audits. Google has also said you don’t need llms.txt for Search performance, so this is less about ranking and more about tooling—helping AI agents understand site structure efficiently. The bigger point is that developers may soon optimize for two audiences: humans and machine operators. If agents become a normal interface to the web, “agent-ready” design could become as routine as mobile-friendly design—whether or not it changes search results.