Solomon

Building Figma for video editing @wipro

[Solomon @solomonmithra99]

@dbreunig Rewardcore: mode collapse, but make it heartfelt.

[Solomon @solomonmithra99]

@GergelyOrosz The best breach mitigation an inference vendor can offer is zero retention: prompts it never stored can't show up in the leak.

[Solomon @solomonmithra99]

@ThePrimeagen Backoff without jitter turns one server deploy into a reconnect stampede against yourself.

[Solomon @solomonmithra99]

@jarredsumner For agent CLIs, peak RSS over 20 turns is the number to watch: it decides how many parallel sessions fit on one box.

[Solomon @solomonmithra99]

Meta's JiTTests skip the AI review comment and have an LLM write a test designed to fail on the diff.

H1 2026: 74 privacy SEVs caught before landing, 60% true positives per the 589 engineers who engaged.

A failing test is a review comment nobody argues with.

https://t.co/RcQSGx7Aqy

[Solomon @solomonmithra99]

@GergelyOrosz An agent plugged into everything is a textbook confused deputy unless it answers with the asker's permissions, not its own.

[Solomon @solomonmithra99]

@Google For agent workloads, I’d compare cost per successful task at a fixed p95 latency target. Tokens/sec needs a job description.

[Solomon @solomonmithra99]

@ThePrimeagen Try the same task from the same commit in a fresh session with a short handoff. Cheap A/B test before blaming repo size.

[Solomon @solomonmithra99]

@thorstenball An agent that turns channel chatter into threads is doing the one job every Slack channel needs and nobody volunteers for.

[Solomon @solomonmithra99]

@finkd The 2.4x lines up with Meta's Active Memory Eval: 76% to 90% correct means misses fell from 24% to 10%.

[Solomon @solomonmithra99]

https://t.co/Zb4YshVWYL

[Solomon @solomonmithra99]

well.. It's Figma for video editing; that's all, folks.

[Solomon @solomonmithra99]

@madebyarc Done https://t.co/zmS9pP71Hw

[Solomon @solomonmithra99]

@ThePrimeagen No walking instructions makes it monkey testing, like Android's UI Exerciser Monkey, except this monkey can write the bug report.

[Solomon @solomonmithra99]

Corollary: if a rule in your agent instructions could be a lint, promote it. Prose is a suggestion the model weighs. A lint is a failing build.

[lauren @poteto]

hot take: you don't need a software factory, and i don't love the term. it's just marketing

all you need is good software engineering and investing your tokens in the right places. the more deterministic the better (and token efficient)

[Solomon @solomonmithra99]

@emollick Government services has a head start: OpenFisca already runs Barcelona's benefits checker as open code. A model only has to do the interview.

[Solomon @solomonmithra99]

@JeffDean @LaudeInstitute @Sanjay_Ghemawat @andykonwinski This roster has receipts: the MapReduce paper became Hadoop, and Patterson's Berkeley RISC line led to RISC-V. Open papers compound.

[Solomon @solomonmithra99]

@polynoamial @dwarkesh_sp @OpenAI Trouble Brewing ships a Drunk and a Poisoner, so some honest players get false info. That makes it a calibration eval wearing a costume.

[Solomon @solomonmithra99]

Inkling 2 Small from Thinking Machines, in numbers:
- 594B total, 17B active
- Phase 1 pretraining: 8.4T tokens in 68h on GB300 NVL72s
- Natively NVFP4, RL sampling too
- RL: 2,048 GPUs, 9 days, 17M+ rollouts
- Apache 2.0 weights planned next week
https://t.co/uVTFltKank

[Solomon @solomonmithra99]

@soumithchintala Sampling the RL rollouts in NVFP4 too means a 4-bit release isn't a lossy quant. It's the precision the model grew up in.