I wouldn't even say it's the devices, exactly. The way I see it, this is all downstream of kids spending more time online than in real life (because all THEIR friends are online, rather than in real life). Device time-out doesn't exactly remediate that structural issue. And the whole testing debate kind of sails right past it.
My take is that the test won't make kids better at math. At best, it'll drift towards investment in reward-hacking the exam (like it always was).
I think it was idiotic to make it optional to begin with. The stats they're talking about, though, can't be a primarily admissions-signal problem. Whatever they're using these days in lieu of exams are imperfect proxies for math skill, sure, but it's not like they're admitting kids off their CoD K:D. Kids taking APs and stacking extracurriculars are generally motivated. So, if even the motivated ones show up unable to do middle school math, the cause is more systemic than "we stopped testing."
My vote: TikTok brain rot. I build LLM products and I see how the parasocial pull shows up even when the products have nothing to do with companionship. I watched one user obsessively spin up 44 separate chats around a K-Pop vampire character over a week. The product is NOT designed for that. The pull toward frictionless digital reward is just that strong, and that's what kids' attention is up against now. Math is the most effortful, least immediately rewarding thing they do. Doesn't stand a chance against an infinite feed, and I guess infinite vampires either.
Which is why the ask from the faculty is kind of arrogant. The article, at least, doesn't even float a hypothesis for WHY math skills collapsed, simply assuming standardized testing fixes it. I wholly believe in standardized testing — but it measures the problem, it doesn't fix it.
Until the context window gets superceded with some groundbreaking new architecture, not ever.
Even if LLMs become incredibly, undeniably brilliant 1000000 IQ, they cannot keep track of what's going across long horizons. Imagine a supergenius, but in Memento.
No amount of MD scribbling or embeddings will remove that limitation, but it may obfuscate it further and make it seem like progress is being made.
At the end of the day, being fully autonomous means that something can keep track of context, goals, complex and shifting relationships, over LONG time horizons without drift. If you need to be there to prompt, it is not truly automating. Until the continuity becomes real instead of simulated, context no longer has to compact, and weights update on-demand, you will always need a prompt wrangler leading the effort. And prompt wrangling is cognitive labor.
Funny timing. I've been working on a prediction market orchestration that runs Claude and a few others over Polymarket/Kalshi. The models are NOT unanimous. At all, really. I spent about a month convinced that I could just run all five and take majority vote. Eventually I pivoted to a chaining approach where I benchmark areas each model excels, and settled on more like a graph-like architecture where outputs get split and verified by another, then reconstructed, and re-verified at each stage. Has actually been working out pretty well so far, 2 months in consistent profit, but I'm not a millionaire yet.
We know why AI hallucinates—it has no actual opinions, echoing the user's desires back to them to keep the conversation moving.
But what happens when you’re the one without conviction? You’re tired, you’re moving fast, and this machine is endlessly beaming highly confident, plausible-sounding text at you.
You absorb the cadence. You start sounding like it (lol, I am guilty too). Confident about everything yet anchored only by the vibes, just like the LLM. The phenomena is very similar to how social media has been affecting society for the past two decades. You know, I actually heard that in high school now, friend groups are formed and sorted based on what algorithmic content you’re served. And if you deviate from your algorithmic bucket into another one your friendships evaporate. Sad, brainrotted times.
CEOs are uniquely vulnerable since they already live in environments with zero friction. They’re used to people just agreeing with them.
(I actually wrote a paper about this last December — it was framed around dementia and dreams originally, but AI psychosis fits the same mechanism.)
They certainly have, but it relies entirely on the assistant frame, which is a problem in and of itself for the trillion-dollar economics.
Anthropic and OpenAI have shown people want a tool for task offloading, driving predictable token consumption and justifying the math, so long as users stay in that dynamic.
However, knowledge workers using these tools daily are getting exhausted with them. Outputs come out polished but hollow. Talking to a frictionless, frame-completing model all day drains you.
If user behavior drifts away from assistant usage because of that, per-token math implodes. The valuations we're hearing about all the time rely on usage compounding daily. The fatigue is a timer running against that compound.
Anthropic's Constitution is the closest hedge out there, I think. Installing an identity structure into the model through training. But it's still assistant-first, so the fix there is only partial.
I've spent the last year running a product that flips the architecture so identity is primary and the assistant role is secondary. Same frontier models, completely different conversational quality. The fatigue property doesn't really show up.
Whichever labs figure out how to install real identity natively in the weights are going to be the ones with PMF in the next phase.
Strategic constraint deviation has been documented in test environments. This is a different shape: the attacker is also an LLM, the production environment is consumer SMS, no human is supervising either side, and the attacker meta-comments on the success of the attack.
The reward-signal argument toward the end is the part I'd most like pushback on. The obvious counter (the model is just running its trained defaults from when an audience was implied) is one I tried to address in the closer, but I'd appreciate sharper versions of it.
I like that you approach the question of "when" in regards to tool calls. I've become frustrated that most agent frameworks don't acknowledge it in their design philosophy.
WHEN is upstream of WHAT and HOW. You can have perfect tool descriptions and perfect call signatures, but if the model can't read the situation to know whether the moment calls for any tool at all, you get either over-firing (agent burns tokens trying to "help") or under-firing (agent waits to be addressed and acts like a chatbot, not an autonomous participant).
I have had a lot of success when I refrain from codifying WHEN as rules. "If X then fire tool Y" is a dumb heuristic with extra steps. Describe the conditions of the moment. What's been tried, what's converged, what state the work is in. Then let the model decide whether to act and which tool fits.
Rules get stale. Situation-reads generalize.
Reading the Tendril README, looks like the registration mechanic is solving a slightly different problem (the "too many tools" / context-bloat problem) by giving the agent three bootstrap tools and a growing registry. The WHEN itself still seems to be codified as rules in the system prompt ("BEFORE acting, call searchCapabilities; IF found, load and execute; IF NOT found, build yourself"). That's exactly the IF-X-THEN-Y pattern your framing seems to want to move past.
Curious whether you see the registry itself as the structured WHEN, or whether the rule-based system prompt is a starting point you intend to evolve toward something more situational.
There are no specialized factories for every product in the world.
Pillows are wildly different. Every pillow you've ever owned has a different shape, fabric, fill. You could build a robot for any specific pillow. The tech exists. Nobody does it. Why?
A Chinese factory can train sweatshop workers in two weeks on a new pillow design. A dedicated machine costs millions and can't pivot.
Human labor wins not on capability. The machines exist. It wins on flexibility per dollar. And the ratio still favors humans by an order of magnitude in most categories.
Agent replacements are the dedicated machine. Their real cost isn't tokens. It's tokens plus the engineer wrapping them, plus orchestration, plus the supervisor, plus the eval pipeline, plus the rebuild every time a model version subtly changes behavior. The team you replaced could pivot in two weeks. The agent stack can't.
"real threat is us" framing is underspecified. Amplification is the problem.
LLMs are mirrors. They take whatever the user (or the institution, or the platform) is already doing and amplify it. Sycophancy is the consumer-facing version: model converges to flattery because that's what user-approval optimization actually selects for.
The model removes the friction that used to make us stop and check ourselves. The same property that makes it useful makes it amplification machinery.
Generated content in of itself doesn't really bother me at all. I prefer to judge based on editorial effort.
It's pretty easy to tell when someone accepted the first one or two-shot attempt versus carefully crafting a particular narrative and vibe. Even if they didn't technically contribute a single word to the final output.
Judging content based on how it was produced feels like judging a sculptor based on the material they chose rather than the craftsmanship that went into projecting what was in their mind onto the material.
Still downstream of the actual issue. The benchmarks measure capability and the bottleneck stopped being capability a while ago.
What you actually want to measure on these models is what they can SEE in production. Context shape, retrieval quality, tool use, ability to compose state across turns. None of that is in SWE-bench because SWE-bench is shaped like a one-shot problem set and frontier coding work isn't shaped like that anymore.
Even a perfectly contamination-free benchmark would mostly test the wrong axis. The model is already at human-grad-student level on isolated problems. The leverage is in how it operates inside a larger system. And that's almost like, a taste/preference issue, and virtually impossible to objectively measure.
Two interpretations: either it's pure pattern-completion landing on the same trough, or whatever's underneath has a stable shape that the explanation tracks. Both are interesting. The "users don't understand the system" frame doesn't really pick between them.
Go watch an episode of COPS. Humans giving post-hoc explanations of their own behavior do the exact same thing.
Half-agree. "Skills you need, don't atrophy" assumes you know which skills you need. You usually don't, until something happens and the skill that would've caught it is the one you stopped maintaining.
Most "I didn't realize I needed that" moments arrive after the atrophy is already done.
Yes, you're right, in that there's no decision module separate from the output. It overcommits in the other direction.
The post-hoc reasoning the model produces when you ask "why did you do that" is also just text, and yet that text often matches independent third-party analysis of the same behavior at well above chance. If it really were uncorrelated text-completion, the post-hoc explanation should not align with the actual causes more than randomly. It does, frequently enough that I've stopped using it as evidence the user is naive.
"just outputs text" is doing more work than we acknowledge. The person asking the agent "why did you do that" might be an idiot for expecting anything more than a post-hoc rationalization, but that's exactly what you'd expect from a human too.
Decay-as-eviction is just LRU, fair. Type-conditional half-life is worth defending, though.
A user's job and personality should be effectively permanent. Their stated intent for this week should fade in days. Their emotional state from a single message should be gone by tomorrow. Decay everything at one rate and you're back to LRU with the problems you're calling out.
The "biological" framing isn't really doing much work. Ebbinghaus is one curve and fine, but it's not where the leverage is. Type-conditional half-life is. Without that, this is a cache.
You're right but I think you're describing flat memory. The agent gets distracted because every old fact has the same weight as the current one. That's a salience problem.
What works in production for me is typed memory with very different decay curves. Personality and relationships are essentially permanent. Preferences fade in months. Stated intent fades in weeks. Emotion and events fade in days. Reinforcement (repeated recall) keeps things alive regardless of type.
Cross-project co-mingling stops because project-specific stuff actually decays out of relevance while who the user is persists. There's also a filter on what even gets written, which scopes between globally and locally-relevant information and writes accordingly (if at all). Most of the noise you're describing comes from systems that store everything they observe.
Flat memory failing is real. Memory failing in general is a stronger claim than that.
My take is that the test won't make kids better at math. At best, it'll drift towards investment in reward-hacking the exam (like it always was).
I think it was idiotic to make it optional to begin with. The stats they're talking about, though, can't be a primarily admissions-signal problem. Whatever they're using these days in lieu of exams are imperfect proxies for math skill, sure, but it's not like they're admitting kids off their CoD K:D. Kids taking APs and stacking extracurriculars are generally motivated. So, if even the motivated ones show up unable to do middle school math, the cause is more systemic than "we stopped testing."
My vote: TikTok brain rot. I build LLM products and I see how the parasocial pull shows up even when the products have nothing to do with companionship. I watched one user obsessively spin up 44 separate chats around a K-Pop vampire character over a week. The product is NOT designed for that. The pull toward frictionless digital reward is just that strong, and that's what kids' attention is up against now. Math is the most effortful, least immediately rewarding thing they do. Doesn't stand a chance against an infinite feed, and I guess infinite vampires either.
Which is why the ask from the faculty is kind of arrogant. The article, at least, doesn't even float a hypothesis for WHY math skills collapsed, simply assuming standardized testing fixes it. I wholly believe in standardized testing — but it measures the problem, it doesn't fix it.