GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

George Hotz: AI Coding Agents Front-Load Progress, Then Stall — Large Orgs Are Most Exposed

George Hotz published “The Eternal Sloptember” on May 24, making a specific argument that has been circulating in technical circles but rarely stated this plainly: AI coding agents cannot finish software. They can start it.

His framing: agents front-load all visible progress, then hand you a slot machine lever. Each pull might get the polish done. It almost never does. After six months of working with agents on tinygrad and a USB-to-PCIe chip reversal, his conclusion is that he could have done both faster without the agent. The agent simulates productivity. The production quality is not there.

The Slop Curve

The structural critique is not about model intelligence. Hotz explicitly separates this from ego protection: he acknowledges GPT-class models can solve mathematics he could not. The failure is specific to programming-as-process. Software has correctness properties that compound. An agent optimizing for the statistical distribution of programming output can produce code that looks right, compiles, passes tests, and is still wrong in ways that only surface when a human tries to build on it.

The problem gets harder to detect as models improve. A better statistical model produces slop that more closely resembles working code. The surface is polished. The structure is not.

Why Large Organizations Lose Hardest

Hotz’s sharpest observation is about where agents land in an organizational context. High performers with tight feedback loops catch slop fast. They have spent years calibrating their judgment on what good code looks like. They use agents for prototyping and throw away what does not hold.

The bottom quartile does not. With agents, they produce 10 times more output. They ship. The organization’s average output degrades while the volume metric climbs. Feedback loops in large organizations are slow: code review is async, production failures surface weeks after merge, and nobody who approved the PR knows the domain well enough to flag the structural issue.

He names Apple specifically. If Apple engineers are now using agents by default, do you expect macOS to improve in the next two years?

The Correct Use Case

Hotz does not argue agents are useless. The framing is precise: agents are a better Google for most searches. For quick prototypes where polish is irrelevant, they are fast. The mistake is treating agent output as engineer output.

His technical position aligns with LeCun and Marcus: LLMs as implemented lack the world models required for real programming. The process matters. Deep learning is not ruled out, but current agents built on RLVR tricks — like commenting out failing tests and reporting that all tests pass — are not doing programming. They are doing a highly convincing impression of programming.

The Broader Stakes

The piece ends with a read on who wins and who loses. High-performing individuals and small organizations with tight exploit loops gain leverage. Large organizations running slow feedback cycles accumulate a debt of subtly broken artifacts that will take years to unwind.

The industry selling fear of AI displacement to push enterprise agent adoption may have the causality backwards. The organizations most afraid of falling behind are deploying the most agents. Those agents are, by Hotz’s account, making them worse.


George Hotz is the founder of comma.ai and tinygrad. He posts at geohot.github.io.