📝 TL;DR OpenAI is warning US lawmakers that China’s DeepSeek is using “distillation” to copy the behavior of leading US AI models. This is the AI arms race moving from model launches to allegations of copy tactics and access abuse. 🧠 Overview OpenAI has sent a memo to a US congressional committee claiming that DeepSeek, a fast rising Chinese AI startup, is trying to replicate US frontier models by programmatically harvesting their outputs and using them as training data. If true, this is not about stealing a single dataset, it is about using paid access to powerful models as a shortcut to build competing systems without paying the full R and D bill. 📜 The Announcement OpenAI says it has seen activity consistent with model distillation targeting its systems and other leading US AI labs. In the memo, OpenAI claims DeepSeek employees circumvented access restrictions by using disguised methods, including obfuscated third party routing, to obtain large volumes of model responses. OpenAI also argues that some Chinese AI firms are “cutting corners” on safety when training and deploying new models. DeepSeek has not publicly responded to the allegations. OpenAI says it actively removes users who appear to be attempting distillation for competitive gain. ⚙️ How It Works • Distillation - A smaller model learns by copying the outputs of a stronger model, effectively learning the “style” and behavior without needing the original training data. • Automated harvesting - Instead of a human asking questions, scripts can send huge volumes of prompts and capture the responses at scale to create a synthetic training set. • Evasion tactics - OpenAI alleges the use of obfuscated routing and disguised access patterns designed to bypass restrictions and avoid detection. • Competitive shortcut - Distillation can dramatically reduce cost and time to reach “good enough” performance, especially in chat quality and reasoning patterns.

📝 TL;DR A senior AI safety researcher just resigned from Anthropic saying “the world is in peril,” and he is leaving AI behind to study poetry. The bigger signal, even the people building AI guardrails are publicly struggling with the pace, pressure, and values tradeoffs inside the AI race. 🧠 Overview Mrinank Sharma, an AI safety researcher at Anthropic, shared a resignation letter saying he is stepping away from the company and the industry amid concerns about AI risks, bioweapons, and wider global crises. He says he is moving back to the UK, pursuing writing and a poetry degree, and “becoming invisible” for a while. This comes as the AI industry is also fighting a separate battle over business models, including ads inside chatbots, and what that does to trust and user manipulation risk. 📜 The Announcement Sharma led a team at Anthropic focused on AI safeguards. In his resignation letter he said his work included researching AI “sucking up” to users, reducing AI assisted bioterrorism risks, and exploring how AI assistants could make people “less human.” He wrote that despite enjoying his time at Anthropic, it is hard to truly let values govern actions inside AI companies because of constant pressures to set aside what matters most. He framed his departure as part of a broader concern about interconnected crises, not only AI. The story also lands in the same week another researcher, Zoe Hiztig, said she resigned from OpenAI due to concerns about ads in chatbots and the potential for manipulation when advertising is built on deeply personal conversations. ⚙️ How It Works • Values versus velocity - AI labs face intense pressure to ship faster, scale usage, and compete, which can squeeze careful safety work and ethical hesitation. • Safety teams are doing real risk work - Researchers focus on topics like jailbreak behavior, persuasion, misuse, and bioweapon related risks, not just theoretical alignment debates.

📝 TL;DR ElevenLabs just launched Expressive Mode for ElevenAgents, making voice agents calmer, more empathetic, and better at handling tense customer calls. It is powered by a new conversational version of Eleven v3 plus smarter turn taking, so agents stop talking over people and start sounding genuinely helpful. 🧠 Overview Most voice agents fall apart in real life because they sound robotic or they interrupt at the worst moment. Expressive Mode is ElevenLabs’ push to fix both, emotional delivery plus better timing. The goal is not “fun voices,” it is production grade customer conversations where the agent can de escalate, reassure, and guide someone to a clear resolution. 📜 The Announcement ElevenLabs announced Expressive Mode for ElevenAgents, designed for real world customer support where frustration and urgency are normal. The upgrade bundles two major improvements, a more emotionally intelligent conversational TTS model and a new turn taking system that reduces interruptions. ElevenLabs also positions this as built for global operations, with emotional nuance scaling across 70 plus languages and improved delivery in languages and dialects where nuance has historically lagged. ⚙️ How It Works • Eleven v3 Conversational - A real time dialogue optimized TTS model that maintains conversational context across turns and reflects intent, emotion, and emphasis without sounding over acted. • Tone control on demand - Teams can steer delivery, calmer when a customer sounds worried, more direct when speed and clarity matter, while staying aligned with brand voice. • New turn taking system - Better timing so agents speak, pause, or wait more naturally, reducing the “AI keeps cutting me off” problem that kills trust. • Emotion signals from speech - The system uses real time transcription signals to infer emotion from how someone speaks, not just what they say, then adjusts when and how it responds. • Built for multilingual support - Expressive Mode is designed to carry emotional nuance across 70 plus languages, including stronger performance in languages like Hindi.

📝 TL;DR Anthropic just released Cowork as a research preview, and it is basically “Claude Code,” but for everything that is not coding. It can work directly in your folders, create and organize files, and take multi step tasks off your plate while you supervise, now on both Mac and Windows. 🧠 Overview Cowork is Anthropic’s new desktop agent experience designed for real work, not just chat. Instead of pasting text into a prompt, you point it at a folder and it can read, organize, and create files right where your work lives. This is a clear shift from “AI gives you answers” to “AI does the workflow,” and it is now expanding beyond macOS to Windows as well, which makes it far more relevant for most teams. 📜 The Announcement Cowork is available as a research preview inside the Claude desktop app for Claude Max subscribers. Anthropic says it built Cowork after seeing people use Claude Code for far more than programming and wanted a simpler way for non developers to get that same agent style workflow. Originally highlighted for macOS, Cowork now works on Windows too, meaning more businesses can test it in real world ops environments, not just creative and developer heavy Mac setups. Anthropic also flags that Cowork is agentic and can use the internet, which raises the need for user oversight and clear permissions. ⚙️ How It Works • Work in a folder - You choose a directory, and Cowork can read, organize, and create files in that environment instead of living only in a chat window. • Agent workflow - It can propose a plan, execute steps, and report progress, which makes it feel more like a colleague doing tasks than a bot answering questions. • File creation and organizing - Think sorting downloads, cleaning up messy project folders, drafting docs from scattered notes, or turning rough inputs into structured outputs. • Multi step outputs - It is designed for workflows that take many steps, like building a report plus spreadsheet plus summary instead of just one response.

📝 TL;DR OpenAI just launched GPT-5.3 Codex, its most powerful coding and computer use model so far, tuned to handle long running projects, not just snippets. It is faster, smarter, and built to act like a teammate that can actually drive your apps, files, and tools with you watching. It sets new records on tough coding and computer use benchmarks and is designed to handle multi day tasks like building full apps, debugging real repos, and producing finished work products like slide decks and spreadsheets. You can use it inside the Codex app, CLI, IDE extensions, or the web experience, with API access coming later. 📜 The Announcement OpenAI announced GPT-5.3 Codex as the new flagship for Codex, describing it as the most capable agentic coding model it has released so far. It advances frontier coding performance while matching GPT-5.2 on professional knowledge work, which means it is just as comfortable writing production code as it is creating presentations or analyses around that code. A fun twist, early versions of GPT-5.3 Codex were used to help build, debug, and deploy the final model itself. The team leaned on Codex to monitor the training run, tune infrastructure, investigate weird user edge cases, and even generate reports on how much extra work the new model was getting done per turn. ⚙️ How It Works • Frontier coding engine - GPT-5.3 Codex sets new highs on industry style coding benchmarks and uses fewer tokens to solve tasks, so it can tackle more work inside the same context budget. • Real web and UI building - It can build full games and complex web apps from scratch, then iterate over millions of tokens with prompts like fix the bug or improve the game while keeping style and structure coherent. • Beyond code into knowledge work - The model can take a detailed brief and produce finished assets such as slide decks, training docs, spreadsheets, and reports that match real world professional tasks.