Corrections
Corrections
When something published here turns out to be wrong, the post is fixed and the change is noted on the post with its date. This page lists every one of those notes. Spotted a mistake? Tell me.
An agentic pentest tool drove the ConnectWise breach; ADK changes telemetry
Dory's model card does report scores per loop count (1, 2 and 4 loops) and recommends 2 by default. The brief first said it didn't. The KV-cache trade-off per loop was added.
Lab, Step 1 (part 3): scale your tiny GPT without fooling yourself
Chinchilla does mention inference savings; the post first said it said nothing about inference. The example model sizes are about 2–3x apart, not 4x.
A Brotli decoder bug reaches Node and Python, and Kimi K3.1 Pro lands
Upgrading is not the only fix for the Brotli bug: until you can upgrade, reject inbound Brotli-encoded requests and stop decoding Brotli on outbound calls (Aikido's advisory).
A Brotli decoder bug reaches Node and Python, and Kimi K3.1 Pro lands
Removed a link to a SWE-Verifier code repository that isn't publicly available.
The resume snapshot is not capped at 2 KB: the README says so, but the code has ignored that limit since April 2026. References to the cap were removed.
One MCP URL for a hundred engineers: what a tool gateway actually buys you
Fixed a broken link to the 20 October research post.
The agent that writes its own workflow: when the plan becomes a program
The multiagent_20261001 type is documented in the multi-agent orchestration docs, which also say workflows are on by default with it. The post first said the version string wasn't in the docs.
Lab, Step 1 (part 2): the training loop is where small GPTs quietly go wrong
Added measured results for the five breakages from the runnable lab: at this model size only removing the mask or the residuals fails clearly; warmup, the attention scale and a 20x learning rate barely change the result. The curves above are illustrative.
Lab, Step 1 (part 2): the training loop is where small GPTs quietly go wrong
Forgetting to divide by the accumulation steps does not multiply the learning rate under AdamW, which normalises gradient size. It does with plain SGD, and it changes what the clipping threshold means. The lesson was corrected.
Meta releases an 816B open-weight model, and Temporal workers scale to zero
SWE-at-Scale's code and models are released (the arXiv page links them). The brief first said the paper didn't commit to releasing code.
Workday opens HR data to your agents over MCP, and Meta ships Muse Wonder
Meta's post names Harvey's Legal Agent Benchmark, FocusBench and SWE-Marathon. It does not name MemoryBench, as first published.
Lab, Step 1 (part 1): build a GPT you can read in one sitting
The starting-loss check (about ln V) catches an over-confident initialisation, not an off-by-one target shift. The post now gives a separate check for the shift.
Wikimedia catches OpenAI agents off-script, and Mistral previews a 1T model
Mistral Large 4 has 52 billion active parameters, not 49 billion as first published.
ReCAP: compacting agent memory from attention you already paid for
ReCAP is not limited to self-hosted models: §5.6 of the paper shows a small open model can supply attention for a hosted-API agent with no significant loss. The post first said hosted-API users couldn't use it.
Agents turn fix PRs into exploits in minutes, and GPT-6.1 Sol undercuts Astra
rclone had about 20 security disclosures in its whole first decade, not 20 a year, so the recent jump is much larger than first stated.