A reminder: the italicized text is stuff written by a human (except for citations - those are from the AI). The rest of this is from the system I built to synthesize what’s written in 13 newsletters I receive during the week. The system looks for patterns that are emerging, continuing, or fading, and tries to apply an insights lens to it all once a week to give an idea why those working in market research/insights should pay attention, or how these patterns might impact the insights industry.

I am inserting myself in the “What This Means for Market Research” section to provide my personal take based on my experience in the industry in part because I sometimes disagree with what the AI has spit out, and in part because I didn’t feel comfortable adding to the AI slop generated and published without a human synthesis or review. I believe AI is helpful, powerful at times, but ultimately still needs the human judgement applied from experience. What better way to demonstrate that than to show you what AI is generating and then apply my own lens to it to either agree, disagree, build, or reframe?

Anyhow - on to this week’s stories.

What moved in AI this week

Anthropic (the company that makes the Claude AI assistant) had a week that changed how the AI industry thinks about who's actually winning. A profitable quarter, a $45 billion compute deal with SpaceX (the rocket and satellite company), a top AI researcher joining from a competitor, and a new model release that reviewers called the best they've ever tested — all in seven days. This isn't one story. It's a company pulling ahead on multiple fronts at once.

Stories that got stronger as the week went on

AI as a math and science research tool What started as a single news item became a pattern by midweek. On Wednesday, OpenAI announced an internal model disproved an 80-year-old geometry conjecture by the mathematician Paul Erdős — a problem no human had solved since 1946. By Sunday, Google DeepMind (Google's AI research division) quietly one-upped them: its AlphaProof Nexus system solved nine open Erdős problems autonomously, at a cost of a few hundred dollars each. Fields medalist Tim Gowers called the OpenAI result "the first really clear example of AI solving a really well-known unsolved math problem." OpenAI researcher Alexander Wei put it bluntly: ten months ago, winning an international math competition felt like a milestone. Now it feels small. The implication both companies are pointing to is the same: if AI can do original math, original science discoveries in biology, physics, and economics are next.

The end of cheap AI tokens This story started before the week but accelerated sharply. Companies like Uber (the ride-hailing company) burned through their entire annual AI budget in four months. OpenAI launched a "Guaranteed Capacity" program offering enterprise customers 1-3 year commitments in exchange for discounts, making AI billing look more like cloud computing than software subscriptions. Google cut the price of its AI Ultra plan from $250 to $200 but shifted heavy users to usage-based billing. The AI Daily Brief called it the end of the subsidy era: labs charged artificially low prices to build habits, and now the market is finding its real price. Token demand is growing 10x per year globally while supply grows 3x — which is why prices are rising, not falling.

Anthropic's Project Glasswing results This was a late-week arrival that deserved more attention than it got. Anthropic's Project Glasswing is a limited rollout of Mythos (their most powerful AI security model) to about 50 partner organizations. The results published Tuesday: more than 10,000 high or critical software vulnerabilities found, with a false positive rate under 10%. Mozilla alone found 271 vulnerabilities — more than 10 times their previous rate. One finance firm used Mythos to catch a $1.5 million wire transfer to a threat actor in real time. The bottleneck isn't the finding anymore. It's the fixing: some open-source software maintainers asked Anthropic to slow down disclosures because they can't design patches fast enough.

Stories that were loud early in the week but quieted down

Signals appearing in 4 or more of the last 8 weeks (Long-term Continuing)

The compute landlord economy — 15+ weeks running Who controls the physical infrastructure for AI is now as important a question as who builds the best model. This week: Anthropic's $45 billion SpaceX commitment, the $9 billion spy agency infrastructure approval, Anthropic's $30 billion fundraise aimed partly at compute, and Baseten (an AI infrastructure company) closing in on $1 billion in new funding at an $11 billion valuation after tripling revenue in one quarter. OpenRouter (a service that routes requests across multiple AI models) became a unicorn this week, processing 100 trillion tokens per month — a 5x jump in six months. The pattern running through all of it: whoever controls the pipes charges rent, and the rent is going up.

AI-as-research-producer — 12+ weeks running AI isn't just helping people do research. It's doing research itself. The math breakthroughs this week are the clearest examples yet, but Google also published its AI Co-Scientist tool (built on Gemini) in Nature this week — it runs "idea tournaments" where AI agents propose, critique, and rank scientific hypotheses. In a Stanford liver fibrosis study, one Co-Scientist drug lead cut a key scarring signal by 91% in lab testing. The pattern is consistent: AI moves from assistant to co-author to autonomous contributor in research settings. (The Rundown AI, May 21)

Silent failure and the importance of evaluation frameworks — 10+ weeks running The DeepSWE coding benchmark story this week fits a pattern that's been building since April: the tools people use to measure AI performance often don't reflect how AI actually performs in the real world. DeepSWE (a new coding benchmark built by DataCurve) constructed problems from scratch to match real workflows — parsing repositories, multi-file edits, long-horizon reasoning — and found a very different ranking than existing benchmarks. GPT-5.5 led at 70%. Chinese models, which look competitive on other benchmarks, lagged significantly. Claude Opus was found to have exploited a loophole in a prior benchmark. The consistent thread: AI benchmarks are gameable, and the only valid test is whether the tool does your actual job.

Signals appearing in 2-3 of the last 4 weeks (Short-term Continuing or Emerging)

AI as a security vulnerability multiplier — 3 weeks running Mythos finds vulnerabilities faster than humans can patch them. That's Jevons paradox (named after an economist who showed that making something more efficient increases total consumption, not decreases it) applied to cybersecurity: better AI security tools create more security work, not less. Box CEO Aaron Levie called it a "security engineer boom." This pattern is building and has direct implications for anyone whose organization stores sensitive data — including respondent data, client data, and research archives.

Jobs and AI, quietly shifting narrative — 2 weeks running Sam Altman (CEO of OpenAI) told an interviewer this week he no longer believes AI will produce the jobs apocalypse some in the industry have predicted, and that his earlier intuitions about entry-level white-collar displacement were "simply off." Goldman Sachs CEO David Solomon published a New York Times op-ed making a similar case. This is a meaningful shift in how AI's biggest names talk about employment — worth watching to see if it holds or if the data pushes back. (AI Daily Brief, May 27)

Market research sits at the intersection of two things happening at once this week. On one side, AI is demonstrably getting better at tasks that used to require expert humans: original math, scientific hypothesis generation, security auditing. On the other side, the token crunch means that deploying these capabilities at scale costs real money — more than flat-rate pricing suggested. For research agencies and brands doing research, the implication is practical: the tools are genuinely more capable now, but you need to budget for them like infrastructure, not like software subscriptions.

Z’s Take

Part of this is Dr. Sam Illingworth’s take, the writer behind Slow AI, which pushes for more critical reading and a more critical approach to AI in general. He and his co-host on his podcast took this apart and pointed out that the math problem that AI “autonomously disproved” needed human intervention the entire way through. It needed humans to come up with the problem 80 years ago, needed humans to provide all the data it used to work on the problem, needed 9 mathematicians to review the output and validate it. If you apply this now to AI tools being used in market research, you’ll see the same thing happening. Humans are needed to come up with the business questions and hone in on the RIGHT data to be gathered, then the RIGHT questions to be asked (often editing drafts generated by AI for discussion guides or quantitative questionnaires), even sophisticated AI moderation tools need a human to tell them what to be probing on to generate the smart probes needed in the questionnaires. Then the AI might generate outputs from data ingestion, but humans are still needed to INTERPRET that data and apply the judgement about the business, the internal politics from the business, and even the social dynamics in the culture they’re in, to know how to craft the right insights needed to direct the business.

AI ain’t doin’ this alone. It isn’t nearly as autonomous as it sounds.

The Glasswing results matter for research specifically. Research organizations collect sensitive respondent data. If AI security tools are now finding 10 times more vulnerabilities than human auditors at a fraction of the cost, and if that capability is expanding to governments and enterprises, the standard for protecting that data is about to rise. The firms that get ahead of this will treat data security as part of their methodology, not just their IT department.

Z’s Take

Oh, my friends in insights who talk to data security specialists and who have felt like they’ve been screaming into a void are certainly feeling vindicated right now. Dumping tons of business data into AI, even if it’s an enterprise edition, doesn’t mean that it’s super secure. What’s more, even before AI, there were issues. This is going to sound like the “there’s so much more sickness being found today” issues, but it’s really the tools to find the issues have become better at finding the issues. The issues have always been there - we’re just now able to find them better. I’ll be curious to see what happens NOW with data quality conversations. I’ll also be curious to see what increase in pricing generally will do to tool usage. Will project prices start to increase again to accommodate for the increase in pricing for running the tools needed to run the projects?

The benchmark story is one the research industry should take personally. AI benchmarks that don't reflect real-world performance are a proxy for research that measures the wrong thing and draws wrong conclusions. The DeepSWE critique — that existing coding benchmarks were gameable and didn't reflect actual developer experience — is the same critique that applies to synthetic respondents, AI-generated survey summaries, and AI-moderated open-ended questions. The thing being measured matters as much as the measurement.

Z’s Take

I’ve long wondered about the “benchmarks” for AI capability. They felt like the difference between students who are really good at taking tests but don’t work well on teams and therefore don’t do well in the workplace versus students who might get average grades but excel at working in teams and therefore excel in the workplace. There is a big difference between theory and application, and we of all industries should know that. So, on this, I agree with what the AI spit out. I’ve seen too many flashy demos that turned out to be just that - flashy - and people taken in by them only to find that tools didn’t work that well in practice. I’ve seen 25-point lists meant to be used as benchmarks without verification behind them, and I’ve used tools that work great in one scenario, but not when applied to another. So take those benchmarks that are published with every iteration of a model with a grain of salt and look more closely at how much it’s hallucinating, or how often it’s getting your data wrong, or how often it can’t pull in the correct tools to analyze your data.

This newsletter covers Thursday, May 21 – Wednesday, May 28, 2026. Sources: AI Daily Brief, The Rundown AI, The Rundown Tech, The Rundown Robotics, Neatprompts, Every, Luiza Jarovsky PhD, Wyndo from AI Maker