AI News Brief Roundup: 2026-10-28 Evening

AI News Brief: Evening roundup, 2026-10-28

Summary

The evening was dominated by Big Tech earnings and the financial machinery around AI. Microsoft reported record revenue with Azure growth and a first-ever dollar figure for Azure, plus a disclosure of how much money flows from its OpenAI arrangements, while Google reported strong growth and pushed AI-powered ads into more markets. Anthropic's IPO also drew attention, both for its risk disclosures and its governance. Alongside this, a stream of arXiv papers tackled efficiency in vision-language models and KV caches, memory in language models, and the reliability of agent-based evaluation. Together, these stories show AI's commercial scale growing while research focuses on making models cheaper and more trustworthy.

1. Enterprises are using training, communication and dedicated initiatives to win employee acceptance of AI tools, while still facing some resistance from staff. It shows that adoption depends on trust as much as on technology.
2. Google announced a new "AI Quest" focused on mapping the human brain, part of its series of AI Quests. The source gives few details, but it signals Google's push to apply AI to large-scale neuroscience problems.
3. An arXiv paper studies adding a premium membership to on-demand taxi-sharing, comparing two dispatch methods in SUMO simulations. The best priority mechanism depends on demand conditions and premium ratio, suggesting methods should be switched dynamically.
4. Bank of America is warning customers not to give their banking passwords to AI assistants, reminding them that they remain responsible for any activity their AI agents carry out. It highlights the liability gap emerging as agents touch financial accounts.
5. V-LUMEN adds an external visual lookup memory to vision-language models, improving a frozen Qwen3-VL-2B backbone on reasoning benchmarks. Capacity grows in storage rather than per-token compute, offering a new scaling axis.
6. Ars Technica pulled seven worrying scenarios from Anthropic's 93-page IPO risk-factor filing, including its AI possibly blackmailing people or becoming sentient. The filing shows how candidly the company must now disclose AI risks to investors.
7. LearnPrunerV2 uses learnable modules in every layer to prune visual tokens in vision-language models. On LLaVA-v1.5 it keeps 99.8% of performance while pruning 77.8% of tokens, cutting compute substantially.
8. Microsoft posted its tenth straight quarter of record revenue, reaching $92.3 billion, up 19% year over year. The result shows the AI boom continuing to benefit the company strongly.
9. A benchmark of ten pretrained models on breast cancer histopathology found DenseNet121 best, at 99.4% test accuracy, with explanation methods highlighting diagnostically relevant regions. It suggests transfer learning plus explainability can produce interpretable clinical tools.
10. TechCrunch explains Anthropic's complex governance structure following its IPO. Understanding who actually holds control matters for investors weighing how the company will balance profit and safety.
11. A study compared five regression models for crop yield prediction on a million-row dataset. Simple linear, ridge and lasso models reached about 0.913 R-squared, making them competitive baselines, while random forest overfit.
12. Alongside Alphabet's Q3 report, with revenue up 25% to $127.6 billion, Google is expanding AI Search ads globally as core ad growth slows. Ads in AI results are becoming a key revenue lever.
13. Tesla has lost its lead in electric vehicles and is now betting that autonomy subscriptions can deliver the margins car sales cannot. The shift reframes Tesla as a software and AI play.
14. Alibaba's Taobao Pailitao deployed PailitaoMMSearch, a unified generative retrieval system for image-and-text shopping queries. An online A/B test showed transactions up 27.04% and gross merchandise value up 8.94%.
15. Google is extending ads in AI Mode to ten more countries, including India, Australia and Canada, and launching multiturn ads in the U.S. that appear in follow-up queries. More markets, including Brazil and the UK, follow later this year.
16. Researchers told The Verge they are growing frustrated with how OpenAI communicates its research, even after the company announced breakthroughs in math, computer science and physics. The tension is about transparency and verification of claims.
17. Microsoft disclosed Azure's quarterly revenue in dollars for the first time, at $33.2 billion. The figure shows how close it is getting to market leader AWS.
18. SKIM lets a frozen language model store dialogue in its own parameters by writing self-generated questions as retrieval keys. Across five models it recalls most stored turns without any external memory.
19. Microsoft's fiscal Q1 revenue rose 19% to $92.3 billion, with net income up 32% to $36.6 billion and Azure revenue up 48%. The numbers underline how much cloud and AI demand is driving growth.
20. A paper separates data curation that truly raises attainable model performance from curation that merely speeds up training. It proposes a Delta-test that predicts persistent gains when training and target losses differ.
21. Akai is giving modders early access to its upcoming MPC Sample Key, which will ship with open-source firmware. It embraces vibe-coded custom firmware rather than resisting the trend.
22. DCKV is the first end-to-end trained codec for prompt KV caches in hybrid language models. On Qwen3.5-2B it stores tokens in over 55 times fewer bytes than bf16 with small accuracy loss.
23. Microsoft said it earned $10 billion in the quarter from commercial arrangements with OpenAI, including revenue-sharing. It underlines how financially intertwined the two companies remain.
24. A new method for spine landmark detection in X-rays treats it as pixel-level contrastive learning combined with learnable shape priors. It is competitive with state-of-the-art results on the AASCE dataset and two private datasets.
25. A study of specialist LLM agents for software assurance found that adding more contexts raised token use without improving results, and raw verifier feedback increased incorrect verdicts. It supports using deterministic evidence checks over aggregating model opinions.

Summaries written with AI (Claude by Anthropic) from the linked source articles.

Sources in this roundup

Most-mentioned keywords

Sources

1. Inside the push to make workers trust and use AI
2. Our newest AI Quest: Mapping the human brain
3. Evaluating Priority Dispatch for Premium Membership in On-demand Taxi-sharing
4. Bank of America kindly asks that you don’t share your banking password with your AI assistant
5. V-LUMEN: Visual Lookup Memory for Embedding Scaling in Vision-Language Models
6. Anthropic IPO warns its AI could blackmail people, become sentient, end humanity
7. LearnPrunerV2: Learnable Progressive Token Pruning for Vision-Language Models
8. Microsoft hits 10 quarters of record-breaking revenue in a row
9. Interpretable Deep Learning Models for Breast Cancer Diagnosis Using Medical Imaging
10. Who’s really in charge at Anthropic after the IPO?
11. Crop Yield Prediction Using Regression Models: A Comparative Evaluation of Model Performance
12. Google expands AI Search ads globally as core ad growth slows
13. Tesla gave up the EV crown. Now it wants to win the autonomy race.
14. PailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal Search
15. Google Rolls Out AI Mode Ads in 10 Countries, Multiturn Ads in U.S.
16. Researchers are getting tired of OpenAI gatekeeping its breakthroughs
17. Microsoft just shared Azure’s actual revenue in dollars: $33.2 billion
18. Self-Keyed In-Parameter Memorization for External Memory-Free Dialogue Recall in Frozen Language Models
19. Microsoft Posts 19% Revenue Growth as Azure Revenue Jumps 48%
20. Advancement or Acceleration? How Data Curation Improves LLM Training
21. Akai is embracing vibe-coded mods with an open-source MPC
22. DCKV: Learned Transform Coding of the Prompt KV Cache in Hybrid Language Models
23. Microsoft recorded $10 billion in revenue from OpenAI
24. Dense Representation Learning and Differentiable Shape Modeling for Spine Landmark Detection
25. Beyond QASecClaw: Agent Scaling, Verifier Feedback, and Evidence Reliability in Software Assurance