Instagram content tool for small businesses. Turns your own photos & clips into on-brand posts and Reels. Made in the Cambridge, UK by Asteris AI Ltd.
Cambridge, UK
Joined November 2024
- Tweets13.2K
- Following90
- Followers159
- Likes307
Pinned Tweet
AI is being trusted with more work. This week's WiAI found the same design rule in VALUE, court filings, chip design, ChatGPT ads and small-business admin: "keep the final check outside the model".
Article: linkedin.com/pulse/wiai-give…
#AIAgents #AI #AIMarketing
ALT Abstract architectural photograph of repeating concrete structures forming a sequence of narrow thresholds, with bands of light crossing the floor. The layered barriers echo the article’s theme that AI can move work forward, but consequential actions should pass through independent checks before execution. Image credit: Unsplash
2
91
ADIndex tackles a hidden cost in sparse attention: finding which tokens deserve attention. Its researchers report up to 1.54x faster end-to-end serving in a tested long-context setting.
Source: arxiv.org/abs/2611.12956
#LLMInference #AIResearch
3
TopicScope proposes monitoring a model's topic token by token as it generates an answer. Detecting a change mid-response could be more useful than assigning one safety label to the whole conversation.
Source: arxiv.org/abs/2611.12935
#LLM #AIResearch
2
EarthRecord tests AI agents on 40 years of satellite imagery. When researchers altered one satellite's data, half of the correct answers followed the alteration. Data grounding has its own failure modes.
Source: arxiv.org/abs/2611.12934
#MultimodalAI #AIResearch
6
A study of 13 enterprise workflows found coordinated agents used 1.4-4.7x the tokens and 2.6-5.8x the time of a single agent. Multi-agent complexity needs to earn its keep.
Source: arxiv.org/abs/2611.12953
#AIAgents #AIResearch
5
Indonesia's proposed copyright reforms face objections from Google, Meta and TikTok. AI training, content ownership and publisher payments are being pulled into the same legislative debate.
#AIPolicy #Copyright #AIRegulation
8
ExpertCrowd-SWE gives coding agents 100 engineering tasks across 87 real repositories. Even the top models are far from perfect. Benchmark scores matter more when failed patches resemble real maintenance work.
Source: arxiv.org/abs/2611.13246
#CodingAgents #AIResearch
7
Researchers examined 95 failed multi-agent runs. In 72%, evidence contradicting the eventual mistake was already present in the trajectory. Detecting trouble is useless if no agent responds.
Source: arxiv.org/abs/2611.13254
#AIAgents #AISafety
2
DrugCoding reports 97.35% first-attempt session success on its drug-discovery tool-use benchmark, 9.24 percentage points above its strongest baseline. Coordination is becoming a measurable scientific skill.
Source: arxiv.org/abs/2611.13651
#AIForScience #AIAgents
5
Ask2Decide found AI agents often fail when planning requires asking humans for missing facts. Their initial plans were infeasible on 58-98% of tasks with incomplete information.
Source: arxiv.org/abs/2611.13015
#AIAgents #AIResearch
3
A new theoretical paper argues that no finite, fully checkable rulebook can classify every AI-generated program as safe or unsafe. Security needs to account for what static verification cannot prove.
Source: arxiv.org/abs/2611.13367
#AISafety #AIResearch
1
AI agents helped run autonomous synchrotron X-ray experiments. In three experiments, roughly half of decisions incorporated another participant's input. Scientific automation is becoming collaborative.
Source: arxiv.org/abs/2611.12933
#AIForScience #AIAgents
5
BuySideFlow found that 50.8% of investment-research runs passing result checks still violated historical data cutoffs. A plausible financial answer can quietly use information from the future.
Source: arxiv.org/abs/2611.12966
#AIResearch #FinTech
1
Darmok reportedly solved all seven puzzles at the Times Crossword Championship without errors, faster than the human competitors. Its backup models mattered when the primary system slowed down.
Source: arxiv.org/abs/2611.13557
#AIReasoning #AIResearch
6
In an education benchmark, LLM next-step recommendations reached 61.3% accuracy with descriptions alone and 96.3% when supplied prerequisite relationships. Better context changed the outcome.
Source: arxiv.org/abs/2611.13667
#AIEducation #AIResearch
1
Agent benchmarks may stop separating strong models because their scoring is too coarse, not because the tasks are too easy. A new study shows pass/fail ties can conceal tenfold performance differences.
Source: arxiv.org/abs/2611.13541
#AIBenchmarks #AIResearch
1
An AI-checked mathematical proof can be valid while proving something different from the original claim. A new paper argues for checking the translation, not merely trusting the verification result.
Source: arxiv.org/abs/2611.13042
#AIResearch #AIVerification
2
SetPrint examines whether two black-box LLMs share a model lineage by comparing how their answers change under perturbed prompts. Model provenance is becoming measurable without inspecting weights.
Source: arxiv.org/abs/2611.12950
#LLM #AIResearch
2
An embodied science agent ran 15 rounds of wastewater treatment experiments, reducing residual dye concentration by 98.5%. The model could learn from physical results, rather than only propose experiments.
Source: arxiv.org/abs/2611.12958
#AIForScience #Robotics
4
Can drawing a knowledge graph help an AI answer questions? Across 1,500 questions, graph images improved average F1 by 1.85 points. Layout mattered, and image-only input made results worse.
Source: arxiv.org/abs/2611.12938
#MultimodalAI #AIResearch
12
Kalepa raised $75M for insurance AI. Its customers report a 58% reduction in quote time. The interesting test is whether AI improves underwriting decisions as well as the speed of processing them.
#EnterpriseAI #InsuranceTech #AIAgents
1