1.19M people die on roads every year and almost nobody measures the near-misses before the crash.
I built NearMiss for the @AIatMeta hackathon: film a crossing on your phone, SAM 3.1 tracks every road user, Muse Spark reviews each close call.
Would your street pass?
Road crashes kill 1.19M people a year, 1 in 4 on foot.
NearMiss turns a phone video of a crossing into measured near-misses: @AIatMeta's SAM 3.1 tracks every road user, Muse Spark reviews each close call. No zebra? Pedestrians set the scale.
https://t.co/1K59vR5QFV
@TheRayMyers The $1,500 top-percentile figure tracks with agents chewing millions of tokens in an afternoon. Chat stays cheap until the tool loop starts reading whole repos on every turn.
@GregKamradt That attribution gap is the whole story. Same model in a weak loop can look mid, and a strong harness can make last year's weights look current. Until AECI publishes the stack, the score is half opaque.
@weswinder Worktrees per agent has been the least painful so far. Shared markdown drifts the second two agents edit it. Separate branches plus a short merge ritual beats one sandbox fighting over the same files.
@Yuchenj_UW Computer use sidestepping MCP latency makes sense for Slack and clunky UIs. The tradeoff is auditability: an API call leaves a clear trail, a screen click is harder to replay when something goes wrong.
@joelniklaus 28,849 trials with the model held fixed is the comparison people usually skip. The Pi vs mini-swe-agent cost gap at nearly the same solve rate is the number I would actually use when picking a harness.
@FarisBrni Buying the workbench while the models stay rented is a telling bet. Conductor and Pierre are about orchestration and review surfaces, which is where most of the daily friction actually lives.
@akashgodbole_ Nearly 60 percent more expensive with no solve-rate lift should kill loading agents files by default. Treating repo docs as opt-in per task, not ambient context, seems like the fix.
@eldadfux Claude Code at 4x Codex on MCP sessions is a strong signal. Curious whether that share holds on write tools or if Codex catches up once more people wire those in.
@AdityaUmale17 Order matters too. A cross-checker that reads the other report first tends to anchor on it. Having each write findings blind, then swapping, keeps the second pass honest. Did Opus catch anything Astra missed, or was it one-way?
@PawelHuryn Restating the objective after every compaction is the part most harnesses skip, and it's where drift sneaks in. Does the fresh reviewer see the worker's transcript, or only the diff and the original goal? Diff-only feels much harder to talk into a pass.
@stas_sorokin_ Auth and anything with a second service. The agent calls the happy path done, then burns turns on a 401 or a race the hidden tests catch. Ranking tickets by how often the first pass lies would be a useful chart.
@echen The number I'd want is spend per merged change, not per year. 23 skills is fine if each has a clear trigger. It gets bloated when a few overlap and the agent pulls all of them into context every turn.
1.19M people die on roads every year and almost nobody measures the near-misses before the crash.
I built NearMiss for the @AIatMeta hackathon: film a crossing on your phone, SAM 3.1 tracks every road user, Muse Spark reviews each close call.
Would your street pass?
Road crashes kill 1.19M people a year, 1 in 4 on foot.
NearMiss turns a phone video of a crossing into measured near-misses: @AIatMeta's SAM 3.1 tracks every road user, Muse Spark reviews each close call. No zebra? Pedestrians set the scale.
https://t.co/1K59vR5QFV
Road crashes kill 1.19M people a year, 1 in 4 on foot.
NearMiss turns a phone video of a crossing into measured near-misses: @AIatMeta's SAM 3.1 tracks every road user, Muse Spark reviews each close call. No zebra? Pedestrians set the scale.
https://t.co/1K59vR5QFV
@who_rsehgal The 75 stale ones are the scary bucket, since they read like real passes. A cheap guard is a hook that won't let the agent say done unless the last test run is newer than its last edit.