A multi-university paper says the benchmarks can be cooked — and the very same week, Meta decides the cure is to hand you the weights and let you check the math yourself. This is the AI Daily Briefing for Wednesday. Today: the Princeton-MIT paper putting every “state of the art” claim under a shadow, Meta’s open-source move through the inference-stack lens, and HubSpot’s 200 updates with no technical report in sight. When an AI company drops a blog post saying their new model is “state of the art” because it topped some benchmark, how do we actually know if that number means anything — or if they just figured out how to ace the test? Yeah, it’s a genuinely important question, and the short answer is: a lot of those numbers don’t mean much anymore. A paper from a multi-university team — Princeton, MIT, Cornell, Brown, and others — published on arXiv puts it bluntly: benchmarking is broken. The problem has two parts. First, saturation. According to analysis from Benchr, most of the benchmarks you see cited most often — MMLU, HumanEval, GSM8K — have scores clustered so high that the top five frontier models are separated by less than three points. A tournament where everyone scores in the upper nineties has, in their words, “stopped sorting.” Second, gaming. Per that same Benchr analysis, HumanEval and MMLU are specifically flagged as benchmarks to retire because they’ve been so extensively trained against that a high score may tell you more about whether a company optimized for the test than whether the model is actually capable. So when you see a headline score, ask: is this benchmark saturated? Was the methodology published, and could a third party reproduce it? And is the company grading its own homework? That last one — companies grading their own homework — what does that actually look like in practice, and why is it a problem? The arXiv paper flags this directly in its title: “Don’t Let AI Be Its Own Judge” — meaning when a model is evaluated using another AI system as the scorer, the incentives get circular and the results get unreliable fast. So watch whether a benchmark score comes with an independent methodology — something a third party could reproduce without access to the company’s internal data. If the only source is the company’s own blog post, with no linked technical report, system card, or external replication, treat the score as marketing copy until someone else can check it. This one's from Gizmodo:
While the new releases will reportedly maintain some proprietary parts for alleged safety purposes, the company apparently plans to open-source the models, likely offering licensing agreements to firms that want to use the model instead of going full black box, like many of its competitors.
Six hundred billion earmarked for AI, and the headline move is open-sourcing the weights. Hard to call that floundering. Feels like Meta just showed us the strategy. And Gizmodo wants you reading the “floundering” word. I’d read it through the inference stack instead: open weights mean nothing if Meta still controls who fine-tunes and who serves at scale. Right, and tie it back to what we just hit on benchmarks — if the yardsticks are bent, shipping open weights is one of the few honest moves left. The outside world gets to run its own evals. If they ship clean. The report already says some parts stay proprietary for “safety.” Always the asterisk. First release under Alexandr Wang, fresh off the Scale AI acquisition. So the data-labeling guy is now running the lab that bought his company to fix its models. No pressure. And the timing lands the same week the Senate killed the federal moratorium ninety-nine to one. You want to compete under fifty different state regimes? Open-source the model and let licensees carry the fine-tuning risk. Very convenient. This one's from HubSpot:
“SMBs don't need more AI hype—they need technology that helps,” said Andy Pitre, EVP of Product at HubSpot. “The products we’re launching in our Spring 2025 Spotlight are helping teams move fast on AI and solve their go-to-market challenges. We've embedded AI throughout our entire platform so businesses of any size can start seeing value immediately, without massive teams or budgets.”
Two hundred-plus features, embedded AI across the whole platform, and the pitch is SMBs feel like they’re falling behind. So the cure for AI overwhelm is... two hundred more things to learn? And notice what comes with this press release. Nothing. No technical report, no eval, no failure rate — just a count of features and a quote from the EVP of Product. Right, and the honest version of this is fine — they’re automating the stuff a sales rep already hated doing. Logging calls, drafting follow-ups. I want to know whether any of those two hundred updates changes what the rep actually has to know, or just kills the busywork. Andy Pitre says SMBs don’t need more hype, they need tech that helps — said in a release announcing two hundred features at once. Pick a lane, Andy. And after the benchmark segment we just did, there’s no number here to even check. At least a frontier lab gives you a cooked benchmark to argue about. This is just a feature count. If AI Daily Briefing is part of your routine, consider subscribing wherever you’re listening. And if you have a moment, leave a quick review — it really helps other people find the show.
We’ve put the links for everything we covered today in the show notes. If something caught your ear, that’s the place to dig in a little deeper.
That’s AI Daily Briefing for today. This is a Lantern Podcast.