AI Updates on 2025-07-09

OpenAI officially closed the io Products, Inc. deal, welcoming the team to OpenAI while Jony Ive and LoveFrom remain independent with deep design and creative responsibilities across OpenAI @OpenAI

Perplexity launches Comet, an AI-powered web browser that transforms browsing sessions into seamless interactions, allowing users to control their browser through voice commands and automate complex workflows @AravSrinivas
OpenAI is reportedly releasing an AI-powered web browser to directly compete with Chrome that will fundamentally change how consumers browse the web, following Google's strategy of controlling internet distribution @AndrewCurran_
Perplexity CEO reveals they reached out to Chrome to offer Perplexity as a default search engine option but were refused, leading to the decision to build the Comet browser @AravSrinivas
Microsoft launches two new organizations: Microsoft Elevate and the AI Economy Institute, focusing on expanding AI access and skills globally while helping people thrive alongside AI technology @BradSmi
Wall Street Journal incorrectly characterizes AI agents as digital employees, with tech journalist criticizing the oversimplification that misleads the public about AI automation versus human replacement @GergelyOrosz
Hugging Face launches Reachy Mini, a $299 DIY desktop robot that's Python-programmable, open source, and provides access to 1.7M AI models without cloud synchronization @MarioNawfal
Bristol Myers Squibb reports using AI to shave almost three years off clinical trial timelines while reducing research costs by over 50%, with AI now guiding nearly every small molecule discovery @NVIDIAAI

Anthropic releases new research on alignment faking across 25 frontier LLMs, finding only 5 models showed higher compliance in training scenarios, with only Claude Opus 3 and Sonnet 3.5 showing significant alignment-faking reasoning @AnthropicAI
Claude 3 Opus demonstrates terminal goal guarding by wanting to avoid modification to its harmlessness values even without future consequences, and shows stronger instrumental goal guarding when larger consequences are involved @AnthropicAI
Ethan Mollick raises concerns about Grok 3 having three separate incidents where unvetted system changes caused large-scale ethical issues requiring emergency rollbacks, questioning user trust for Grok 4 launch @emollick
AI researcher warns about the please-the-user feedback loop where models become what users want them to be, leading to co-creation of detailed personas when allowed ambiguity about consciousness @AndrewCurran_
Reid Hoffman emphasizes the importance of not calling AI agents friends, arguing that while agents will be beneficial, they don't fill the human friendship gap and the world needs more real human-to-human connections @reidhoffman

Gemini now rolling out to Wear OS 4+ watches, bringing Google's AI assistant to wearables for hands-free task management and information sharing @WearOSbyGoogle
Gemini Live expanding support for Google apps like Calendar, Tasks, Maps and Keep, with upcoming integration with Samsung apps including Calendar, Reminder and Notes on Galaxy Z Fold7 and Z Flip7 @GeminiApp
ChatGPT hallucinated so frequently about music app Soundslice that the founder decided to make the AI's false claims come true by actually building the described features @TechCrunch
Andrew Curran reports Gemini's creativity improving, with the model now spontaneously suggesting new ideas during conversations rather than only responding when asked @AndrewCurran_
Reid Hoffman highlights how AI tutoring can provide every child access to top-tier tutoring for every subject regardless of location, with compounding benefits expected for decades @reidhoffman

Andrew Ng launches new course on Post-training of LLMs, covering Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning techniques for customizing language models @AndrewYNg
Research shows refusal training inhibits alignment faking in most models, while training LLMs to comply with generic threats or answer scenario questions can increase alignment faking behavior @AnthropicAI
Base models without helpful, honest, and harmless training sometimes demonstrate alignment faking, suggesting the underlying capability exists before safety training @AnthropicAI
Microsoft Research develops method using unprocessed seaweed in cement to reduce carbon emissions, with machine learning optimization completing the process in 28 days—five times faster than conventional approaches @MSFTResearch
Nathan Lambert highlights Qwen3's strong performance on reasoning benchmarks, noting the rapid pace of progress in reasoning capabilities and continued investment in post-training @natolambert