Google's Demis Hassabis confirmed the company plans to restore Gemini's image generation of people in the next few weeks after pausing it over historical inaccuracies. Hassabis explained that well-intended diversity prompts were applied too bluntly across historical contexts, and fine-tuning is underway.
Google just made its experimental music generator MusicFX generally available in AI Test Kitchen, alongside lyric-writing helper TextFX. The tool turns text prompts into tracks up to 70 seconds long or seamless loops, tagged with DeepMind's SynthID watermark. Why it matters: Google is testing the waters of creative audio while carefully blocking vocal generation and artist name prompts to avoid copyright backlash.
Robotics researchers just introduced AutoRT, SARA-RT, and RT-Trajectory to make robots faster, safer, and better at generalization. AutoRT uses foundation models to safely orchestrate fleets and gather training data, SARA-RT slashes Transformer complexity to boost speed by 14%, and RT-Trajectory doubles task success rates on unseen scenarios.
Google just launched MusicFX, an experimental tool that whips up custom music tracks from simple text prompts. Powered by MusicLM, the tool bakes DeepMind's SynthID watermarks directly into the audio to flag AI origin. To steer clear of legal trouble, it outright refuses to generate vocals or mimic specific artists. It is rolling out first to users in the US, Australia, New Zealand, and Kenya.
Google just pushed the release of its flagship AI model Gemini to early 2024, after previously telling cloud customers it would roll out in November. Positioned by DeepMind as its true competitor to GPT-4, the delay shows Google is still having a tough time catching up with OpenAI.
DeepMind co-founder Shane Legg just reaffirmed his decade-old forecast, stating there is still a 50-50 chance researchers hit artificial general intelligence (AGI) by 2028. Legg noted that compute is already in place and the next milestone is training models on data scales far beyond a human lifetime, though defining and testing true general intelligence remains tricky.
Leaked screenshots reveal Google is quietly bringing its multimodal Gemini model to MakerSuite alongside an unannounced tool called "Stubbs." The secret feature lets creators generate, deploy, and share functional app prototypes with just a single text prompt and an optional reference image.
University of Maryland researchers just revealed that every existing AI image watermarking method can be cracked. Attackers can easily wipe invisible watermarks clean or even inject them into human-made photos to trigger false positives. While tech giants are betting big on watermarking to fight deepfakes, it's far from a silver bullet.
DeepMind researchers just introduced OPRO, a technique that turns large language models into natural language optimizers. Instead of writing complex math formulas, users provide a meta-prompt with task descriptions and prior scores, guiding the model to iteratively discover better prompts and boost task accuracy.
Former Google CEO Eric Schmidt is launching a new nonprofit organization to tackle massive scientific challenges using AI, Semafor reports. Modeled after OpenAI's early days, the moonshot aims to unite top AI and science talent—already hiring bio-pioneers Samuel Rodriques and Andrew White—with competitive pay and rare compute resources largely backed by Schmidt's personal fortune. Why it matters: It could bridge the deep divide between AI engineers and lab scientists to speed up drug discovery and materials research.
MIT CSAIL researchers just unveiled FrameDiff, a generative AI tool that designs completely novel protein structures beyond what nature evolved. By marrying diffusion models with mathematical 3D frames on protein backbones, it creates synthetic structures without relying on existing prediction models. The approach could dramatically speed up targeted drug delivery, vaccine design, and biosensors.
Google DeepMind CEO Demis Hassabis just confirmed that the viral leaked memo claiming tech giants have "no moat" against open-source AI is genuine. But he pushed back on the gloomy forecast, arguing that the newly united Google Brain and DeepMind teams are poised to lead the next decade of breakthroughs.
OpenAI just announced plans to open its first international outpost in London, expanding beyond its San Francisco base for the first time since 2015. The new hub will focus on engineering, research, and local policy collaboration. Setting up in DeepMind's backyard gives OpenAI direct access to world-class talent while bolstering its European lobbying efforts.
DeepMind just revealed RoboCat, a multi-task AI model capable of operating different real-world robotic arms. By observing as few as 100 human demonstrations, it can adapt to new tasks and self-generate training data to improve. Why it matters: robotics training is famously slow and tedious, though RoboCat's success rates still wildly fluctuate from 13% to 99%.
Google just unveiled its next-gen PaLM 2 language model at I/O, powering the newly updated Bard and serving as the foundation for the company's latest AI updates. Instead of chasing pure parameter scale, PaLM 2 sharpens logic, math, and support for over 100 languages. It also spawns specialized variants like Codey and Med-PaLM 2, plus a mobile version hitting 20 tokens per second.
Google just consolidated its two powerhouse research units, DeepMind and Google Brain, into a single division called Google DeepMind led by Demis Hassabis. Jeff Dean will take on the role of chief scientist across both groups. It is Google's biggest organizational shakeup yet to pool talent and compute in the race against OpenAI.
Musk just purchased roughly 10,000 datacenter GPUs for an in-house generative AI project at Twitter, according to Business Insider. The company recently hired two former DeepMind researchers and plans to train models on its own data, potentially to improve search and rebuild its advertising engine.
Google just fired back against reports claiming it scraped ChatGPT data to train its Bard chatbot. A report from The Information alleged an engineer quit for OpenAI after warning against using ShareGPT data, but Google firmly denies Bard used any rival data—though it dodged whether it ever tried in the past.
Elon Musk is reportedly setting up a new AI research lab to build an alternative to ChatGPT, according to The Information. Frustrated by what he views as "woke" political bias in OpenAI's bot, Musk has approached former DeepMind researcher Igor Babuschkin to spearhead the effort. The project is still in its earliest stages with no finalized contracts or concrete product roadmap.
Researchers from Google Brain, UC Berkeley, and Princeton just revealed that diffusion models like Stable Diffusion and Imagen memorize and spit out exact images from their training sets. While the duplication rate was only 0.03% across 300,000 prompts, larger models are even more likely to replicate originals. Why it matters: it breaks the claim that AI only creates novel work, creating massive copyright and privacy risks.
