LLMs are lagging indicators
Why you should hire recent graduates
I’m an AI optimist but I also recognize that large language models (LLMs) are fundamentally more backward looking than predictive. Sure they can be engines of innovation but they’re best used as archives of consensus. The key concept is “temporal misalignment,” the phenomenon where performance degrades when they encounter facts, slang, or cultural shifts that happened after their training data was frozen. Good research documents this limitation (e.g. the 2025 study Is Your LLM Outdated? A Deep Look at Temporal Generalization). There is a “nostalgia bias” in foundation models, where systems default to historical data and resist newer, less statistically probable information.
A recent Nature study found that AI-augmented science also tends to look backward not forward, looking primarily at established fields, with markedly less focus on edge topics.
An LLM can only “learn” a new concept when it appears often enough in the training corpus (Common Crawl, Reddit, digitized books) to achieve statistical heft. The time until it does is a latency period. So new teen slang or a new trend has to move from Discord or TikTok comments to more static and stable sites on the internet before an LLM can see it. The same way kids know a meme is dead when their parents have heard of it, by the time AI “knows” it, it’s old news. LLMs, like Boomers, are rarely at the cutting edge of a cultural phenomenon. It’s the same for professional knowledge: by the time an industry practice is documented enough to train an LLM, innovators may have evolved past it.
I’ve written elsewhere about the “last mile” problem, where human judgment refines the data to offer more granular and specific analysis. I’m writing here about the first mile problem. In logistics, the last mile is expensive because delivery becomes idiosyncratic at the doorstep: each package goes to a different address with different access requirements and different recipient schedules.
The first mile is expensive because pickup is unknowable and unpredictable. A sender might not even have thought about sending a package yet. Only a few thoughts turn into physical packages to be packaged, addressed, and brought to the post office. The first mile is the gap between “I’m thinking of sending a package” and “it’s in the system.”
The future does not arrive with a tracking number. A first mile advantage is the ability to be in the right community, in the right room, in the right conversation, to recognize that a signal is worth capturing before it has a name.
What does this mean for hiring?
In economic forecasting, a leading indicator, such as the volume of new building permits or manufacturers’ new orders, signals a trend before it becomes obvious in the broader economy. A lagging indicator, such as the unemployment rate or corporate earnings, confirms a pattern after it has already occurred. Lagging indicators are useful because they validate what you see; they provide certainty for making policy and drawing up contracts. Leading indicators are useful for looking ahead, for planning. They offer opportunity for preemptive action and competitive maneuvering.
LLMs generate lagging, not leading, indicators. They validate and optimize established consensus. They are like the Consumer Price Index, tracking the value of concepts the culture has already purchased. For a business owner under pressure to jump on the AI bandwagon, thinking about leading or lagging helps clarify how best to use AI. LLMs are great for tasks involving stable data and stable rubrics. LLMs are great at drafting standard contracts, summarizing historical reports, or coding a known function. They can confirm trends and discover past patterns and trends. It captures the “beta”—the average market performance—with the most fidelity.
However, relying on LLMs for forward-looking strategy is risky. A firm should not depend on an algorithmic system for identifying emerging markets or cultural shifts. In fast-moving sectors like fashion or digital media, novelty will always be ahead of any foundation model. LLMs are not ideal for first mile prediction or for last mile specificity.
How should new graduates leverage this fact?
Both firms and college graduates should understand first- and last-mile needs. An energetic, market-aware recent college graduate should pitch understanding of emergent sentiment (“vibe”) or pre-consensus intuition as a leading indicator. Young people live on the edges of networks where data is raw and contradictory. They read the signals of activity that precede the economic contracts.
Young people also “get” AI—many have been using it for years, even if they’ve mostly been told they should know how to prompt, how to “humanize” output to fool teachers, how to code. While they generally possess higher technical proficiency than their elders, their greater value lies in their function as a sensor network. They will be ahead of what LLMs know.
When hiring managers interview a candidate, they should ask: “what in your life did you notice before anyone else did?” Their answer is worth more than any prompt engineering skills.
Smart firm owners and hiring managers in the AI era should see the labor market as divided between people who can analyze past and present data and people who look to the future while understanding the present. Yes, invest in AI. But don’t short the future by over-indexing on the past.
Smart college graduates should view their temporal immediacy as their primary asset. Capitalize on this latency. Know what LLMs can do and know what they can’t. The LLM is the archive; you are the source of renewal.
Most hiring managers think about skills. They should be thinking about time. AI is unmatched at compressing the past into usable form. But the past is cheap and getting cheaper. The scarce resource is contact with the future, being present when a signal is still noise, when a pattern has no name yet. Firms that understand this will hire for talent accordingly. They’ll stop asking candidates to demonstrate what a model can already do and start asking what they’ve seen that no model could have told them.
LLMs may have knowledge cutoffs and be 'out of date', but this is not an intrinsic limitation of the technology; just how they are trained and deployed right now. They can train in realtime on all new data before almost any humans have seen it.
It is *convenient* for AI labs to do big batch scrapes, single big training runs, and have intense vetting and redteaming of a specific checkpoint which gets dropped in a big bang several months later, and it is not particularly valuable to them to know teen slang or memes in real time; but this approach is not intrinsic to LLMs. There is no reason that LLMs could not know kid slang even before 99% of kids know it, if the social media companies or AI labs wanted to.
Many ML technologies are deployed and trained in realtime (Chinese e-commerce and social media are especially good at this), or at tempos like hourly or daily. This is especially common in fast-moving or adversarial contexts like recommenders or spam filters, where even enormous models may be retrained or trained from scratch constantly, like Tiktok or Google Ads. And LLMs can be too, new text is jut more tokens to predict...
One way to think of it: an LLM *must* "train faster than realtime" because otherwise it could not catch up on centuries of written text in a mere few months of training. If an LLM read and train on everything written in the past, say, 182,000 days in just 182 days (500 years vs 6 months), then it must go through an average of 182,000/182 = 1000 text-days every training-day, or to put it another way, 1000 text-minutes every training-minute.
Clearly it would not be difficult to do just a little more training on what text happened to be written today, and so contains today's new slang. (In the past I've estimated that you could probably keep a frontier LLM up to date on all new high quality English text in realtime by running at most a few hundred GPUs 24/7 - hardly anything!) So set up appropriately, an LLM totally could know emerging slang within minutes, long before almost anyone knew it. Do a minibatch of a few million tokens every couple of seconds, replicate it across the inference datacenters worldwide over the private backbone links over the next minute, switch over the live traffic, and done. Now the LLM knows the "fnargle" slang that some Philadelphia kids invented in their Tiktok posted a minute ago and which a few million kids will see overnight - but it knew that slang hours or days before they all did.
(And given the low latency speed of hardware like Cerebras chips, ever-increasing power of small dense models, and within-datacenter/backbone networking speeds, it might be possible to do a full update loop in the time it takes to do a single keystroke! So since many apps log drafts or keystrokes, all the AIs worldwide could well know what that kid in Philadelphia has said before he has even said it. It is a distressing fact about our world that you can do a *lot* with AI in the time it takes Substack to render a 'reply' button in your browser. And censors are well-aware of the inhuman speed of AI knowledge dissemination and how it can outpace mere human communication, eg. https://ai.meta.com/blog/harmful-content-can-evolve-quickly-our-new-ai-system-adapts-to-tackle-it/ )
Such speed is just one of the many ways in which AI are superhuman. ("There's plenty of room at the bottom", down where the milliseconds slowly inch by like snails with quadrillions of aggregate operations happening in total in parallel... https://gwern.net/blog/2025/llms-can-be-faster https://gwern.net/doc/ai/scaling/hardware/2018-sandberg.pdf https://gwern.net/note/faster )
The leading/lagging indicator framework is genuinely useful, though I'd push on one thing: continuous retraining on real-time data is already here at meaningful scale with Perplexity, Grok live search, and web-enabled Claude — which means the 'dead meme when parents know it' effect has a shorter half-life than two years ago. The harder version of your thesis is that even with real-time data, LLMs are trained to reproduce statistical consensus, so they'll find trending signals faster but still systematically underrepresent outlier and pre-consensus ideas. Is there a class of 'first mile' problems where the signal is structured enough that AI has the advantage over human intuition, or is the human edge specifically in pattern-breaking rather than pattern-finding?