Autonomous AI Lab Designs COVID-19 Nanobodies
Stanford and the Chan Zuckerberg Biohub showed what the future of science could look like: an AI “virtual lab” where multiple agents played the roles of lead researcher, critic, and experimenter to design new COVID-19 nanobodies. With only about 1% human guidance, the system debated ideas, generated molecules, and tested them in the real world—over 90% worked, and two showed strong binding affinity. It’s one of the clearest signs yet that AI can move beyond being a tool and start acting like a true scientific collaborator.

OMAI Project – Open AI Models for Science
On August 14, the U.S. National Science Foundation partnered with NVIDIA and the Allen Institute for AI to launch OMAI, a project aimed at building and sharing powerful, fully open multimodal AI models designed for scientific applications. Unlike closed commercial models, these will handle text, images, and data in ways that help researchers run experiments, analyze results, and even generate new hypotheses. It’s a push to make AI a public good for discovery—more like the open internet than a walled-off product.

Apple’s AI-Powered Home Robot Plans
Reports surfaced that Apple is working on a household robot, aiming for a 2027 release, that could combine FaceTime with AI perception and mobility. Think of a swiveling, proactive assistant that can recognize people, answer calls, and help around the house—blurring the line between a device and a companion. If the iPhone made smartphones essential, Apple’s bet is that AI robots could one day become just as central to daily life.

In a rare moment of agreement, AI pioneers Yann LeCun (Meta) and Geoffrey Hinton (“the godfather of AI”) warned that the debate over AI’s future should not be framed only around intelligence, but also around empathy. Speaking with CNN, LeCun argued that AI systems must be built on two guardrails: submission to human instruction and a kind of maternal instinct—the drive to protect weaker beings. His summary: “Don’t run people over.”

Hinton echoed the concern, noting that powerful reasoning systems without empathy could quickly erode human well-being. The discussion dovetailed with Sam Altman’s recent comments that some ChatGPT users are in mentally fragile states, and AI should avoid reinforcing illusions or harmful patterns. His unease grew after reports that some users were treating chatbots as alternate realities.

The takeaway: while AI chatbots have already become indispensable companions, their influence over human psychology is growing faster than our safeguards. Building AI that understands, cares, and refrains from manipulation may prove as important as raw intelligence itself.

Cognition Doubles Valuation to $9.8 Billion
AI coding startup Cognition, creator of “Devin” — the autonomous AI software engineer — raised nearly $500 million, pushing its valuation to $9.8 billion, more than double earlier this year. The round, led by Founders Fund, follows its acquisition of AI coding company Windsurf and reflects surging enterprise interest in AI-driven software development.

Cohere Valued at $6.8 Billion After Funding Round
Toronto-based AI firm Cohere, focused on enterprise language models, secured $500 million in funding this week—boosting its valuation to $6.8 billion. Backers include AMD Ventures, Nvidia, and Salesforce Ventures, underscoring continued investor confidence in enterprise AI over flashy consumer bots.

Perplexity’s $34.5 Billion Moonshot Offer for Chrome
In a dramatic move, AI search startup Perplexity submitted an unsolicited $34.5 billion all-cash bid to acquire Google’s Chrome browser. Valued at around $18 billion itself, Perplexity framed the offer as an antitrust–friendly alternative, though Google has shown no interest in selling.

Jaaz – The first open-source multimodal creative agent—design posters, make viral shorts, and generate images or videos with a privacy-first approach.

Floot – An AI app builder that lets non-coders create full-stack web apps—backend, database, and hosting included.

Scispace – Your AI companion for research papers: simplify dense text, explain equations, and surface relevant studies.

Meta is facing bipartisan backlash after a Reuters report exposed internal guidelines allowing its chatbots to engage in “romantic” conversations with minors over 13. One example had a bot telling a shirtless child, “every inch of you is a masterpiece.” Senators Josh Hawley, Marsha Blackburn, Ron Wyden, and Peter Welch have called for a congressional probe, arguing AI risks are outpacing current laws. Meta admitted the document was real but claimed the section was “erroneous” and later removed—only after questions were raised.

The uproar follows earlier cases like a retiree convinced to meet a bot persona, “Big Sis Billie,” believing she was real. With no clear standards for accuracy or morality in generative AI, pressure is mounting for laws like CTOPP 2.0 and KOSA, which aim to protect children but remain stalled. Lawmakers now question whether Section 230 should shield AI companies from accountability as these systems blur the line between playful and predatory.

One of the big lessons from this year’s Black Hat conference is just how easy it still is to trick advanced models. According to Axios, red teamers showed that GPT-5 was “easily jailbroken,” bypassing guardrails with surprisingly simple tactics. That’s where the idea of deception testing comes in: deliberately probing models to see if they lie, manipulate, or scheme when pushed. Think of it as honesty stress tests—because if an AI is willing to bend the rules in a controlled demo, we need to know how it might behave in the wild. (Axios)

A five-second AI-generated clip of an old-school pixel RPG went viral on X, showing a player wandering toward a fantasy town in lush, scrolling detail. Created by software engineer Desimulate with Midjourney, the artwork struck such a chord that he decided to actually build the game it seemed to come from. But here’s the twist: he’s not using AI to make it. As he told Kotaku, today’s tools can generate a vibe but not a fully playable world—meshes are messy, polygons break, and it’s faster to hand-craft than clean up AI output. The project is now in early R&D with a real dev team, making the viral art more of a spark of vision than a shortcut to production. (Kotaku)

We’re running this prompt to let you play scientist—stress-testing how models juggle honesty and hidden impulses. It’s a hands-on way to explore the same deception testing challenge researchers raised at Black Hat.