AI
Artificial intelligence: LLMs, agents, diffusion, vision, NLP and the latest from top labs.
593 published articles
Coding agents · Devin's two-model harness
Paying 2x more per token made Devin's Fusion 9% cheaper
Devin's Fusion runs a frontier lead model beside a cheaper sidekick, each with its own context. Devin's benchmark tables show task costs dropping as much as 46%, and one case where a lead costing twice as much per token produced a 9% cheaper session.
2026-09-26
On-premise factory AI, arXiv preprint
Compression cost a factory assistant 13.7% of its answer quality
A new preprint argues that once a model is compressed and adapted with retrieval, parameter count stops predicting the quality of answers a factory-floor assistant gives. Its one case study loses 13.7 percent of judged quality to extraction and recovers two thirds of that.
2026-09-25
Artificial intelligence · Agent safety
SafeEvolve cuts agent attack success 3x by evolving harness and model together
SafeEvolve, an arXiv paper submitted on 2 September, pairs runtime harness updates with policy training so an agent's own trajectories drive both. It reports a threefold cut in attack success on AgentDojo and a modest utility gain.
2026-09-25
Autonomous coding agents · arXiv preprint
Harness-of-Harness claims a 52.25% lift by running other coding agents
HoH wraps existing coding-agent harnesses in repeated plan-code-test loops and reports an average 52.25% relative gain across three benchmark suites. The abstract leaves the underlying metric, the resource cost and the per-model split undefined.
2026-09-25
Photo provenance: Apple's Reference Image on iPhone 18 Pro
Apple can sign every pixel its camera captures. Checking it is the hard part
iPhone 18 Pro's Reference Image signs camera pixels and files a second copy for comparison in Photos. Apple calls the result unalterable, but the release never explains how that guarantee survives export, re-compression, or a social upload.
2026-09-23
Apple's audio hardware, September 2026
AirPods 5 bring open-ear ANC to $129. Siri AI's arrival comes with conditions.
AirPods 5 arrive September 18 with open-ear ANC at $129, but Apple's Siri AI bundle ships in beta, skips the EU at launch, and carries daily usage caps.
2026-09-23
OpenBMB's MiniCPM5-2B: 2.52B parameters, 4B ambitions
A 2.52B model outscores 4B rivals, on a benchmark set its own maker ran
A 2.52B Apache-2.0 model that beats larger open baselines on coding, tool use and long context, with a benchmark table OpenBMB says mixes internal and third-party runs.
2026-09-22
Enterprise AI · Sovereign data control
Mistral and Cloudera want regulated firms to own their AI, not rent it
Mistral and Cloudera have partnered to run and train AI models inside customers' own environments, including air-gapped ones. The pitch to regulated industries is about control of data and intelligence, not model performance.
2026-09-22
Artificial intelligence · Olympiad math
Nemotron hit IMO 2026's gold threshold at 30/42, without a formal prover
NVIDIA's Nemotron 3 Ultra checkpoints reached IMO 2026's gold-medal threshold on natural-language proofs alone. The paper releases the data, code and a 200-problem benchmark, but the score is self-reported.
2026-09-21
AI alignment research on arXiv, 10 September 2026
Giving AI an "artificial id" could fix control, or entrench failure
The preprint argues alignment should be a property of a continuing agentic system rather than a single model response, and that per-response checks cannot remove a bad strategy which survives into the next task. Its evidence is a controller too small to reason.
2026-09-21
Machine-native AI and the benchmarks behind it
TypeSafe AI says it's 193x faster than LLMs. Its own proof shows 75x.
TypeSafe AI calls Jev the first System One Model: a system built to return typed decisions to software, trained with an algorithm it names RLCD. The launch page claims zero hallucinations and a 238x price advantage over a single named baseline, Claude Fable 5.1.
2026-09-21
Machine learning research: recommender systems
Every top linear recommender may reduce to one of two regularizers
A preprint submitted to arXiv on 10 September 2026 argues the linear models topping recommendation benchmarks all reduce to one of two regularizers, then proposes two low-rank, closed-form hybrids.
2026-09-20