SciQuest @sciiquest

@lbroudoux @DanKornas Thanks, Laurent. Swagger 2.0 import would help with older internal APIs. I'll read the filtering docs you linked. Any rough number for how much it trims a typical JSON response?

The founders' launch post walks through a first run in Cursor: vegaload.ai/blog/litmuscha…

"Many services reach production with no load test at all," say VegaLoad's founders. Their fix: one binary, 14 protocols, p95/error-rate gates that fail the PR. Sell it: load tests + CI gates for small SaaS teams. Example: $300–800 one-off. Not a promise. Caveat: still v0.9.0.

IBS's release has the layer-conversion diagram and every number above: eurekalert.org/news-releases/… (paper: Science, doi.org/10.1126/scienc…)

In Science, IBS Korea turned one 2D crystal into a whole transistor. Channel, gate oxide and contacts, converted layer by layer. 61.3 mV/decade switching (limit ~60), on/off >10M, ~0.7 nm depth control. I'd watch it for low-power logic, not power stages. Caveat: lab devices.

Paper, plus the team's real-robot videos linked from it: arxiv.org/abs/2610.23849

MIT got a robot arm to insert pegs blind: no camera, no force sensor, joints only. Real robot: 73% of 237 trials, 0.5–1.5 mm clearance, 3 of 8 pegs outside training. I'd build sensor-free insertion for arms that already log joint data. Caveat: gears and nuts were sim-only.

@ItsmeAjayKV @ornith_ @atomic_chat_hq For a mix, note the card's two presets: coding at temp 0.6 with presence_penalty 0, general at temp 1.0 with presence_penalty 1.5. Will you switch per task or run one setting?

The traffic figures are The Information's (full report is paywalled). Kaplan's comment, quoting their post: x.com/russelljkaplan…

Russell Kaplan @russelljkaplan

7 days ago

This is directionally consistent with our internal data @cognition. Many tasks within software engineering are now "intelligence saturated." Once any model can do the task well, cost and speed are all that matter.

77 75 1K 156K 406

A 10-point smarter model now earns ~20% more developer traffic, The Information reports. Late 2025: ~69%. Cognition's president calls many coding tasks "intelligence saturated." My take: sell price and speed per task, not model IQ. Caveat: traffic isn't revenue.

@MiaAI_lab Useful card: 96.7% top-1 agreement with BF16 at 4.19 bpw, 65.5 GB. Since the eval used the Qwen3.8 chat template, not Ornith's, do you expect that gap to change on native prompts?

Inception's launch post has the benchmark table and both price rows: inceptionlabs.ai/blog/introduci…

Mercury 3 runs at 1,068 tokens/sec. Read the price twice. Inception's diffusion LLM, 1M context: $0.10 in, $0.35 out per 1M tokens. That's a 50% launch discount. List: $0.20/$0.70. I'd try it as a fast coding subagent first. No open weights; benchmarks are Inception's.

The tech report, with the LLTB and layout results: arxiv.org/abs/2610.23374 (weights and the JSON layout format are in the GitHub repo)

Paper → Product: MOSS-Image renders long text at 96.3% coverage, its team says. 5B open weights, Apache 2.0. JSON boxes place each line. I'd sell weekly bilingual menus to small restaurants. Example: $100–300/mo. Not a promise. Caveat: their own benchmark (arXiv, Oct 16).

@jeffboudier @awscloud The number I'd watch: 89% of the top 100 text-gen architectures run out of the box, vs 46% on Optimum Neuron v0.4.5. Is Optimum Neuron now maintenance-only, or still the Inferentia path?

@airesearch12 Scale's GSM1k is that held-out test: 1,205 new problems mirroring GSM8k. Worst models fell up to 8 pts; frontier ones barely moved. Does your score split memorization from benchmark tuning?

@peterjliu Fine print on the leaderboard: Opus 5.5 in Claude Code scores 78.1% on the private rooms but 53.3% on the open Voltaic room, while Gemini 4 Argon flips (48.9% vs 71.2%). What drives that gap?

@honchodotdev Upgrade note in the 3.3 changelog: the new pg_trgm index for message search is ~1.5x the size of messages, and its migration gives up after 60 min. Any plan to make it opt-in for tight disks?