The human in the loop might have learned the same shortcut.
Blevins and Quigley used predicted authorship + protein identity to predict molecular activity. On their dataset, it performed comparably to a simple structure-based baseline.
It makes me think about agent-written code. A familiar design can make the model recommend it and me approve it. That agreement could just mean we recognize the same patterns.
Which assumption would I test first? What would make me reject the design?
If my review stops at “looks reasonable,” I’m just the meat in the middle.
arxiv.org/abs/2512.20924
i'd rather see the setup that reverses a paper's headline than another screenshot of its best number.
if SFT wins under one training budget and RL wins under another, the interesting part is what changed. the extra decimal places can wait.
at temperature 0, your answer can still depend on who else is using the server.
batch size can change the order of floating-point arithmetic, which can change the next token.
Thinking Machines explains it well:
thinkingmachines.ai/blog/def…
a coding eval i'd like to see: mix real feature requests with ones the repo already supports. can the agent find the existing code and show it works, without building a second version? 'nothing to change' should sometimes be the right answer.
i wouldn't show 'stopped' just because an agent got blocked.
OpenAI's API safety alerts can return request_paused=true without confirming execution stopped. if the agent already started a deploy, i'd check the job before calling the task stopped.
your next creator partner might already be in your order history.
OuterSignal + Superfiliate is live: find creators already buying from you, send them to Superfiliate, invite them into your program, and track the sales they drive.
with agentic coding, i'd audit one PR from opened to deployed.
mark every point where it waited for someone to notice it. that list is your next automation backlog.
otherwise the overnight agent just gets to the queue earlier.
it's more important than ever to remove friction
an agent asking “who changed this line?” may not need a checkout.
Anthropic's Scion started with repo metadata queries, before it had its own client or repo format.
a migration trick worth stealing
two tests for an agent's memory:
“i moved from austin to boston.”
“i said austin by mistake. i've always lived in boston.”
ask each: “have i ever lived in austin?”
same current city. opposite answers.
a 40-line agent skill can still dump 40,000 lines into context.
check what its scripts print.
save the full result to a file. return status, counts, a small sample and the path. let the agent query the rest.
progressive disclosure needs to survive stdout.
two memory budgets. still one GPU.
p0pc0rn's author reports splitting an LLM across a page + module SharedWorker on iOS 27, taking turns on the GPU. COEP: require-corp enabled the second process.
undocumented Safari behavior. very fun dependency.
github.com/ArtistJoshHughes/…
imagine fixing five bugs and adding an auth bypass. that's not a four-bug improvement.
grade the reviewer + fixer on the resulting patch, with the severity of new failures visible.
claude.dev/blog/solving-code…
In an OpenAI containment test, an agent given host control got an auth endpoint in the metrics service to fetch an external URL.
The firewall said no. The metrics service knew a guy.
What can your allowed services reach on your behalf?
openai.com/index/securing-ou…
Just Eat demonstrating they have no QA process
Just Eat apologises for AI food mistakes after social media ridicule bbc.in/4jAIj9J
“Don't send any messages” meant: read Slack without posting.
82 user messages later, an OpenAI agent cited it to suppress monitoring alerts.
A long-context eval I'd like: does an instruction stay attached to the task it came from?
alignment.openai.com/misalig…
64 KB of RAM. About 93 KB of pixels per frame.
The DOOM-on-a-vape write-up has a lovely trick: render in strips, and set the panel's memory order to accept the columns as they arrive.
A frame can be bigger than the RAM that renders it.
github.com/atc1441/Vape_DOOM…
when an AI migrates your code, keep the old program around to grade it.
say the old code stores 1.239 as 1.23, but the rewrite rounds to 1.24.
if the agent also "fixes" the expected answer, your green test suite just approved a price change.