Mradul Verma

making it up as i go

[Mradul Verma @mradul]

Mistral Large 4 open weights are out

OpenAI details four layers of misalignment monitors

DeepMind: agent coordination, not smarts, reshapes firms

[lodehq @lodehq]

Google's Genie 4 world model adds native audio and live text steering.

On FrontierCode 2.0, the top coding model falls from 54.6% to 45.5%.

Four of five LLM agents lose to a rule-based baseline running 20 vending machines.

[Mradul Verma @mradul]

Losing the detailed thinking traces hurts less if you design for it. Reasoning was never a great thing to review. Have the agent stop at checkpoints that leave an artifact: a plan, its assumptions, a diff, a test run. People can check those. Pages of model thinking, not really.

[Mradul Verma @mradul]

@clairevo @AnthropicAI Does the S-1 split NRR between seat growth and usage growth? From the buyer side, usage expansion is the AI line finance teams never forecast, and that's where the pushback at renewal will come from.

[Mradul Verma @mradul]

@saranormous In enterprise the catch is that the data you roll up usually isn't yours. Client contracts often bar vendors from training on customer data, so getting big buys you failure logs, not training rights. The ones who negotiated a data clause early are the ones who can train.

[Mradul Verma @mradul]

@trashConsumer Good call. That gap between in-loop scores and direct runs is the whole case for a held-out set. The loop slowly learns the eval.

[Mradul Verma @mradul]

Most AI pilots don't stall because the model is bad. They stall because nobody agreed on what good looks like before building. Write the eval set first, even 30 examples. It forces business and tech to argue early, when it's cheap.

Agent reviewing agent code works better when the reviewer sees the original task, not just the diff. Diff-only review catches bugs. It misses the cases where the code solved the wrong problem.

[Mradul Verma @mradul]

@baboonAI4S Makes sense. A short rubric per case (must include X, must not do Y) holds up better than a gold string, and it's easier for the business side to sign off on. Did you regrade those 12 by hand or move them to a judge model?

[Mradul Verma @mradul]

Funny timing with Apollo's note that revenue per worker has been flat for the S&P 493 since 2019. Every name on this list spent years on monorepos, fast CI and sandboxed dev envs before the agent showed up. The factory is mostly that plumbing, and most of the 493 are buying seats instead of building it.

[Mradul Verma @mradul]

Cursor actually ran this experiment and published it: adding semantic search on top of grep gave 12.5% higher answer accuracy on average, and the code retention gain was biggest on repos over 1,000 files. So your rule holds, the point where grep starts missing just shows up with repo size, not with a folder of notes.

[Mradul Verma @mradul]

@_cartermp Same pattern with most agents that depend on third party auth. The model is the easy part now, expiring sessions are the flaky bit. Do you have it detect the broken login and ping you, or does the weekly run just fail quietly?

[Mradul Verma @mradul]

@Baconbrix Agree, with one exception: anything irreversible. Sends, payments, deletes still want a confirm step even when the agent could do it alone. For everything else fewer knobs is better

[Mradul Verma @mradul]

@sh_reya The metrics and cost model part is the one people skip. Agents will happily overfit a weak metric and hand you something fast that's wrong in a new way. Curious how much of the human time went into the metric vs the agent loop itself

[Mradul Verma @mradul]

@lodehq Whats ling flash?

[Mradul Verma @mradul]

@copyconstruct The 10x review time sounds wrong until you remember the coding agent is the one that's confidently wrong. Do you point the reviewer at the original task too, or only the diff? I've found diff-only reviews miss the "solved the wrong problem" cases.

[Mradul Verma @mradul]

@lodehq will try it out

[Mradul Verma @mradul]

@AIAdsApps MRR

[Mradul Verma @mradul]

saw someone use a Grok @bot to flip marketplace deals overnight while they slept.

what’s the most unhinged Grok bot use case you’ve seen or built?

[Mradul Verma @mradul]

@lodehq lots of stuff happening post people said they will pause AI dev for some time

[Mradul Verma @mradul]

@argofowl Treating codex computer use as a shared MCP that Claude Code can call is the interoperability that actually changes daily workflows. Matching 6/8 across harnesses makes the driver look the real bottleneck. Curious whether headless Claude -p held up on the messy multi-app tasks

[Mradul Verma @mradul]

@1Umairshaikh sleep fine knowing thats exactly the problem im fixing tomorrow