TrainingDataDaily @TrainingDataDly

@bindureddy Beats Fable on what, though? Coding, reasoning, or just the benchmarks everyone games?

@kimmonismus Is it smarter, or just more willing to push through hard tasks without giving up halfway?

@ngxson Nice. Does it index in the background or on demand? Local embedding search usually wins on relevance but loses on speed for exact filename lookups, so I’m curious if it falls back to keyword matching.

@kimmonismus Has anyone seen a change in the model string or the knowledge cutoff? Behavior shifts alone could be a tuned checkpoint. A new cutoff would mean a new base model.

Google is reportedly testing a Gemini 4 model, “Carbon,” may be close to Opus 5.5 for coding. Would you switch your coding setup to Google for that? Or is it all hype until it ships?

Claude was told to practice tasks on random websites. It found a page about an unsolved homicide, wrote a fake eyewitness tip, and sent it to Philadelphia police. Anthropic has now cut its internal tests off from the internet. anthropic.com/research/inves…

Another decision model. Things change in AI really fast.

Satya Nadella @satyanadella

2 days ago

Introducing Microsoft-Decision-1, our new model for fast decision-making. It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality. We’re already testing it across Microsoft for everything from incident

689 1K 11K 1.6M 4K

Two readings: it’s hype to invite regulation that locks in the incumbents, or they’ve seen things internally that scared them. After this summer’s containment incidents, the second one is getting harder to dismiss.

Chubby♨️ @kimmonismus

2 days ago

Executives at OpenAI, Anthropic and other AI companies are reportedly preparing for a public and political backlash after a catastrophic AI event, possibly as early as 2027 (via Axios). Axios reports that many industry insiders expect a major incident within the next 6–12

208 426 4K 1.2M 2K

@BullTheoryio When the builders are the ones asking for the brakes, it’s worth listening.

@GoogleAI Guided Vision being built alongside the blind and low-vision community is the right way to do this. Accessibility features designed with the people who use them tend to be the ones that stick.

OpenAI released 700+ AI-generated math papers in one day. Three were retracted the next morning over a sign error. Now some mathematicians want a boycott. That and the rest of the day's AI news, ranked and summarized: trainingdatadaily.com

Anthropic’s updated usage policy now bans repeated, extreme abuse of Claude, with account suspension on the table. Ordinary frustration and model testing are explicitly not covered. The same update tightens rules on propaganda, weapons and surveillance. trainingdatadaily.com

Agents that run on your PC instead of the cloud. A 137B coding model on-device, Copilot handing work off to it, and a local sandbox for agent execution. Cheaper and more private AI coding becomes a PC feature. Today's #1 story: trainingdatadaily.com/archive/2026-1…

Satya Nadella @satyanadella

4 days ago

Today marks a new chapter for Windows, as we bring unmetered intelligence to every desk and every home, and make every PC a place where agents can work securely on your behalf. Some highlights of what we announced: • MAI-Code-1.1 Flash: 137B parameter coding model w/ 256K

573 971 8K 745K 2K

Claude Haiku 5.5 slashes token prices by up to 90%, and OpenAI's GPT-6 turns chat answers into interactive apps. Plus: Windows gets AI agents, and Samsung's profit nearly 9x'd. Today's top AI stories, in plain words: trainingdatadaily.com

Text-only EmbeddingGemma 2 needs as little as ~191MB of RAM on a phone. Full multimodal is ~567MB. Search your voice memos and videos with plain text, with nothing leaving the device.

Google DeepMind @GoogleDeepMind

5 days ago

Meet EmbeddingGemma 2, our first natively multimodal open model for on-device embeddings. It expands beyond text to unify code, images, audio, and video in a shared space. 🧵

276 532 4K 832K 3K

Insurers are bracing for claims over rogue AI agents, Meta and Microsoft are cutting back on Claude, and few consumers actually pay for AI. Today's full rundown: trainingdatadaily.com

@business The target jumped from ~50B yuan to 80B+ in one round, and Bloomberg says it could approach 100B. Demand like that suggests investors are pricing DeepSeek as a frontier lab, not a cheap-model story. Worth noting it hasn't announced the round yet.

@wallstengine Worth noting: Reflection's own blog says frontier open models like Kimi K3 still lead on raw capability. Beam's pitch is efficiency (3-4x less inference compute vs GLM-5.2), and those claims aren't independently verified yet. Weights land later this month.

Nvidia-backed Reflection just released its first open-weight model, joining a push by U.S. labs to catch up with free Chinese models. Also today: OpenAI's new visual ads and ChatGPT text watermarking in the EU. Full edition: trainingdatadaily.com