Andrej Karpathy posted today about leaving an AI “autoresearch” agent running on his nanochat project for roughly two days. The agent ran about 700 experiments, found multiple real issues that he had missed, and improved the “Time to GPT-2” leaderboard from 2.02 hours to 1.80 hours. That is an 11% improvement from an automated research loop, not from a bigger team or a longer planning meeting.
The details matter because the fixes were not cosmetic. Karpathy said the agent surfaced problems like a missing scaler on QKnorm, missing regularization on Value Embeddings, misconfigured banded attention, and wrong AdamW betas. It also found improvements that transferred to larger models. His broader point was blunt: frontier labs will all do this. AI agents will do more of the R&D work, and humans will contribute on the edges.
If you run a small business, the headline is not “model training got faster.” The headline is that the AI vendors you rely on are about to improve their products at a much faster pace, because AI is starting to accelerate the research process itself.
Why this matters outside the lab
Most SMB owners do not care about QKnorm. They care about whether the tools they pay for get more useful, more reliable, and more affordable.
Karpathy’s experiment points to a shift in all three.
Until now, improving AI systems meant assembling scarce researchers, testing ideas slowly, and relying on human judgment to decide what to try next. That process does not disappear, but it changes when an AI agent can propose experiments, run them in sequence, evaluate results, and keep searching while the humans sleep.
That means product teams can learn faster. Faster learning turns into faster model improvements. Faster model improvements turn into faster changes in the software layer that small businesses actually touch: chat assistants, sales copilots, proposal generators, support agents, scheduling tools, and industry-specific AI software.
The practical effect is simple. The gap between “good enough last quarter” and “obsolete next quarter” gets wider.
The new risk: buying AI like it is normal software
A lot of SMB software buying still follows an old pattern. Pick a vendor, compare a few feature lists, sign a one-year contract, and revisit the category next budget cycle.
That approach is getting dangerous in AI.
When the underlying systems can improve every few weeks, and when vendors can use agents to speed up model research, the pace of product change stops looking like SaaS and starts looking more like a moving target. Some tools will compound quickly because they sit on top of fast-improving models and teams that know how to integrate new capabilities. Others will stall because they are basically wrapping last season’s model with a nice interface.
For small businesses, this creates a strategic problem. The wrong AI stack can lock you into tools that look polished but improve too slowly. The right stack can give you a steady stream of capability gains without forcing a full system replacement every six months.
Four practical takeaways for SMBs
1. Choose vendors that improve in public
Do not just ask what a tool does today. Ask how fast it has improved over the last six months.
Look for a visible shipping rhythm. Product changelogs. model upgrades. new workflow features. better reliability. stronger admin controls. If a vendor cannot show a pattern of meaningful improvement, assume they may fall behind as the market accelerates.
In a slower market, product polish could carry a vendor for a year. In this market, shipping velocity matters more.
2. Avoid long lock-in unless the ROI is already proven
This is a bad moment to sign rigid annual commitments for immature AI products. The category is moving too fast.
Prefer monthly terms, pilot programs, or limited-scope deployments where possible. If you do commit annually, make sure the workflow is already valuable enough that future model improvements are a bonus, not the entire business case.
Do not pay for roadmap promises. Pay for current utility.
3. Build workflows, not tool dependence
The companies that adapt fastest are the ones that know the workflow they want, not just the brand they like.
For example, if your sales team needs help turning call notes into follow-up proposals, define that workflow clearly: inputs, approval steps, systems touched, output quality, and turnaround time. Once that workflow is defined, you can swap tools more easily when a better option appears.
If your process only exists inside one vendor’s magic prompt box, switching gets painful. If your process is documented, tool changes become manageable.
4. Review your AI stack more often
An annual software review is too slow for AI.
For your highest-impact use cases, review the market quarterly. That does not mean ripping out tools every 90 days. It means checking whether your current vendor is still competitive on quality, cost, security, and speed of improvement.
This is especially important for customer-facing use cases like chat, support, outbound communication, and internal knowledge retrieval. These are the areas where better models and better agent systems can produce immediate business gains.
What this means for planning in 2026
Karpathy’s post is one datapoint, but it fits a larger pattern. AI is moving from “humans use AI tools” to “AI helps build better AI tools.” Once that loop starts working, progress speeds up.
For SMBs, that means planning should change in three ways.
First, budget for experimentation. If AI tools improve faster, you need room to test better options without blowing up your operating plan.
Second, prioritize flexibility over perfect standardization. Standardizing too early on a weak vendor can cost more than tolerating some short-term tool churn.
Third, treat AI capability as a competitive variable, not a background utility. If your competitors can adopt faster, better tools every quarter while you are locked into something stale, that gap shows up in response times, staff efficiency, and customer experience.
Small businesses do not need to run 700 experiments on transformer internals. But they do need to understand what happens when the companies building AI start using AI to do more of the research work.
The pace picks up. Product rankings change faster. Late decisions get more expensive.
That is the real takeaway from Karpathy’s experiment. The benchmark win is interesting. The market implication is bigger: the AI tools small businesses depend on are about to evolve on a much shorter cycle, and the businesses that stay flexible will capture the upside.
Source: Andrej Karpathy on X