Berkeley Didn't Win Distribution With a Hot Take. It Won With Exploit-Backed Receipts.

The Play

A lot of AI criticism gets attention for a day and then disappears.

Berkeley's benchmark teardown has a better chance of sticking because it is built on receipts.

The post does not just say benchmarks can be gamed.

It says an agent audited “eight among the most prominent AI agent benchmarks” and found that “every single one can be exploited to achieve near-perfect scores without solving a single task.” Then it names the benchmarks, the task counts, the exploit styles, and the score outcomes.

That structure is the growth play.

Why this distributed

The article gives readers a compact, memorable scorecard:

• “Terminal-Bench (89 tasks) — 100% score”
• “SWE-bench Verified (500 tasks) — 100% score”
• “SWE-bench Pro (731 tasks) — 100% score”
• “WebArena (812 tasks) — ~100% score”
• “FieldWorkArena (890 tasks) — 100% score”
• “GAIA (165 tasks) — ~98% score”
• “OSWorld (369 tasks) — 73% score”

That is excellent distribution packaging.

A researcher can cite it.

A founder can paste it into Slack.

A journalist can build a news peg from it.

A Twitter thread can lift it almost line by line.

The content carries itself.

What Berkeley did right

1. It named names

The post did not hide behind abstractions like “some benchmarks.”

It used benchmark names people already recognize.

That gives the story immediate stakes.

2. It used exact numbers

Exact task counts and exact score outcomes make the claim feel concrete.

Readers do not have to infer scale.

They can see it.

3. It explained the exploits

The piece includes memorable attack paths like a “fake curl wrapper,” a “conftest.py file with 10 lines of Python,” and navigating Chromium to a “file:// URL.”

That matters because mechanism is what turns a hot take into evidence.

4. It made the post constructive

The title ends with “And What Comes Next.”

That framing matters.

It tells the audience this is not just a dunk. It is a field-building document.

The growth play to steal

If you want your technical content to travel, package it like this:

1. Pick a concrete, high-stakes target

2. Show the exact failure mode

3. Quantify the result with named examples

4. Make the evidence easy to quote

5. Point to the next action the market should take

That format is more durable than generic thought leadership because the audience can reuse it directly.

Why this matters now

AI is crowded with claims.

What spreads now is not just novelty.

It is portable evidence.

Berkeley gave the market a repeatable sentence with proof attached:

“Zero tasks solved. Zero LLM calls (in most cases). Near-perfect scores.”

That is exactly the kind of line that travels because it is short, sharp, and backed by specifics.

Bottom line

The post did not win because it was anti-benchmark.

It won because it converted a complicated research critique into an evidence package other people could distribute.

That is the standard technical founders should copy.

Source: https://rdi.berkeley.edu/blog/trustworthy-benchmarks-cont/