Last Week in AI #330 - Groq->Nvidia , ChatGPT Apps, US AI Genesis Mission
Nvidia buying AI chip startup Groq’s assets for about $20 billion in largest deal on record, OpenAI opens ChatGPT to third-party apps via its Platform, and more!
Nvidia buying AI chip startup Groq’s assets for about $20 billion in largest deal on record
Nvidia agreed to license Groq’s inference technology and acquire essentially all of its assets for about $20 billion in cash—its largest deal to date—according to Disruptive CEO Alex Davis, a major Groq investor. Groq, valued at $6.9 billion after raising $750 million three months ago, said it entered a non-exclusive licensing agreement with Nvidia and that CEO Jonathan Ross, President Sunny Madra, and other senior leaders will join Nvidia.
In a note to employees, Nvidia CEO Jensen Huang said the plan is to integrate Groq’s low-latency processors into the NVIDIA AI factory architecture to broaden support for inference and real-time workloads. Nvidia CFO Colette Kress declined comment, and Huang emphasized Nvidia is licensing Groq’s IP and hiring talent, not acquiring the company itself. Groq will continue operating independently under CFO-turned-CEO Simon Edwards, with GroqCloud excluded from the transaction and continuing without interruption. Davis said Nvidia is receiving all of Groq’s assets aside from the nascent cloud business, effectively transferring the core chip IP and hardware.
OpenAI opens ChatGPT to third-party apps via its Platform
OpenAI is now accepting submissions for third-party apps inside ChatGPT, creating an app directory where approved tools are discoverable and usable directly in the ChatGPT interface. Developers can submit apps built with OpenAI models through the developer platform; listings undergo automated and manual review for policy compliance, safety, and technical reliability before approval. Once approved, apps appear alongside built-in tools and can be invoked in conversational workflows without external installs. The rollout is global, and eligibility depends on adherence to usage policies and quality guidelines, transitioning the prior ad hoc GPT sharing toward a structured marketplace.
Amazon has a new leader for its ‘AGI’ group as it plays catch-up on AI
Amazon CEO Andy Jassy said Rohit Prasad, head of the company’s AGI efforts and longtime Alexa leader, will leave next year as Amazon hits what he called an “inflection point” in AI. Peter DeSantis, an SVP at AWS with deep infrastructure experience, will now lead a new division spanning advanced AI models, custom chip development, and quantum computing. The reshuffle puts renowned AI researcher Pieter Abbeel in charge of the frontier model research team, while DeSantis unifies Amazon’s AI stack from models to hardware.
Microsoft, Google Join US AI Genesis Mission to Advance Scientific Discovery
Two dozen major AI companies, including OpenAI, Microsoft, Nvidia, Amazon Web Services, and Google, have joined the Trump administration's "Genesis Mission," a federal initiative aimed at accelerating scientific discovery and energy innovation through artificial intelligence. Announced via executive order last month, the program will coordinate research across government agencies and leverage the Energy Department's national laboratories' computing resources to automate experiment design, speed up simulations, and generate predictive models for breakthroughs in energy, manufacturing, and drug discovery.
John Carreyrou and other authors bring new lawsuit against six major AI companies
Related:
A new lawsuit filed by authors including John Carreyrou targets Anthropic, Google, OpenAI, Meta, xAI, and Perplexity, alleging the companies trained large language models on pirated copies of their books. The complaint argues that prior relief—namely the Anthropic $1.5 billion settlement offering roughly $3,000 per eligible writer—fails to address the “massive willful infringement” of using stolen books to build models that generate billions in revenue. This follows a judge’s earlier ruling that while training on pirated copies may be legal under current interpretations, the underlying act of book piracy is not, leaving a gap authors want courts to address directly. The plaintiffs say the proposed Anthropic settlement “serves the AI companies, not creators,” and seek accountability specifically for training on datasets sourced from infringing repositories.
In a parallel suit, Adobe faces a proposed class action alleging it trained its SlimLM small language model on SlimPajama-627B, which the complaint says is a derivative of RedPajama that includes the Books3 corpus of 191,000 books. Plaintiff Elizabeth Lyon claims her guidebooks appeared in a processed subset of SlimPajama, arguing Adobe relied on a deduplicated, multi-corpora dataset that nonetheless incorporated Books3 via RedPajama copying and manipulation. Books3 and RedPajama have repeatedly surfaced in litigation, including cases naming Apple Intelligence and Salesforce for alleged training on copyrighted works “without consent and without credit or compensation.”
Other News
Tools
Google launches Gemini 3 Flash, makes it the default model in the Gemini app. Google’s new Flash model is cheaper and faster than its predecessor, matches or exceeds top models on several benchmarks, and is rolling out as the default in the Gemini app and via Vertex AI/enterprise APIs, with previews for developers and enterprises.
OpenAI’s GPT-5.2-Codex advances software engineering with better reasoning and context understanding. It builds upon the capabilities of GPT-5.2, adding improvements in context compaction, large code refactoring, Windows environment performance and cybersecurity.
Z.AI launches GLM-4.7, new SOTA open-source model for coding. Available via Z.ai’s Open Platform and APIs, GLM-4.7 expands context length, improves reasoning, coding, and multimodal (text+vision) performance.
Anthropic makes agent Skills an open standard. They’re a new method for developing specialized AI agents using files and folders. Those folders include instructions, resources and scripts that Claude and other LLMs can leverage to perform specific tasks.
Klarna Launches Open Standard for Agentic Commerce. A new protocol offers a live, standardized feed covering 100 million products and 400 million prices across 12 markets, with an API compatible with Google Merchant, Shopify, Facebook Catalog, and CSV/JSON to make merchant inventories discoverable by AI agents.
Meta introduces new SAM AI able to isolate and edit audio. The system supports text, time-segment, and visual prompts to separate or remove specific sounds, is available to download and via Meta’s Segment Anything Playground, and has limitations when isolating very similar audio events and on taking audio as a direct prompt.
ChatGPT launches a year-end review like Spotify Wrapped. The feature, available in several English-speaking markets to eligible free, Plus, and Pro users with certain memory and history settings enabled, delivers personalized awards plus a poem and image summarizing usage; it’s optional, privacy-focused, and unavailable to Team, Enterprise, or Education accounts.
Mozilla’s new CEO says AI is coming to Firefox, but will remain a choice. Mozilla plans optional AI features in Firefox while diversifying revenue beyond search and expanding the browser into a broader ecosystem of trusted software.
Business
Chinese robotaxis due in London next year as Lyft and Uber reveal tie-ups. Trials of Baidu’s Apollo Go RT6 driverless electric taxis are set to begin in London in 2026, prompting national security warnings about potential Chinese technological dependence and data risks.
Waymo resumes service in San Francisco after robotaxis stall during blackout. A citywide outage caused many Waymo vehicles to stall at intersections as systems treated dark signals as four-way stops, leading to a temporary suspension while the company assessed the incident and began applying lessons learned as service resumed.
OpenAI is reportedly trying to raise $100B at an $830B valuation. The cash would cover rising compute and development costs as OpenAI expands model releases, global deals, and developer tooling while navigating competition, chip supply limits, and fundraising scrutiny.
How China built its ‘Manhattan Project’ to rival the West in AI chips. A prototype EUV lithography machine—reportedly reverse-engineered by ex-ASML engineers in a secure Shenzhen facility and completed in early 2025—is now under testing as part of a state-backed effort to close China’s advanced chipmaking gap.
Vibe-coding startup Lovable raises $330M at a $6.6B valuation. Funding will support deeper third-party integrations, expanded enterprise features, and infrastructure—databases, payments, and hosting—to enable full-fledged application development on the platform.
Yann LeCun confirms his new ‘world model’ startup, reportedly seeks $5B+ valuation. The company, Advanced Machine Intelligence (AMI), is chaired by LeCun with Alex LeBrun as CEO and aims to build “world model” AI that simulates cause-and-effect to reduce LLM hallucinations; it’s reportedly seeking roughly €500M at a €3B valuation.
ChatGPT’s mobile app hits new milestone of $3B in consumer spending. Most of the $3 billion arrived in 2025 as consumer spending—largely from Plus and Pro subscriptions—surged, outpacing revenue growth timelines of major apps like TikTok, Disney+, and HBO Max.
Cursor continues acquisition spree with Graphite deal. The deal, reportedly for well above Graphite’s $290M valuation, brings stacked pull request and AI-driven code-review tools into Cursor’s stack to reduce buggy AI-generated code and speed developers from draft to ship.
Research
Emergence of Human to Robot Transfer in Vision-Language-Action Models. Scaling up robotic foundation models and robot training data enables them to leverage egocentric human videos without explicit alignment, yielding about a 2x improvement on robot tasks where robot data is scarce.
T5Gemma 2: Seeing, Reading, and Understanding Longer. Lightweight encoder-decoder LLMs (270M–4B) pretrained on ~2T tokens with multimodal vision inputs, tied embeddings, merged attention, and extended positional interpolation handle long contexts up to 128K while matching or exceeding comparable decoder-only baselines.
LLaDA2.0: Scaling Up Diffusion Language Models to 100B. The authors convert pretrained autoregressive checkpoints into diffusion-style masked reconstruction models using two-stage continual pretraining and a post-training pipeline to produce instruction-tuned 16B and 100B models.
Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models. Scaling a sequential, domain-by-domain cascaded reinforcement learning pipeline on 8B/14B Qwen3 bases yields unified thinking/non-thinking models that match or exceed specialized reasoning and code models across many benchmarks, with open training recipes and released models.
Universal Reasoning Model. The study demonstrates that recurring, depth-wise nonlinear computation and modest architectural tweaks (short convolutions and truncated backpropagation) make parameter‑efficient Universal Transformers substantially better at multi-step abstract reasoning tasks like ARC‑AGI and Sudoku, achieving higher pass@1 accuracy and faster convergence under matched budgets.
VL-JEPA: Joint Embedding Predictive Architecture for Vision-language. A joint image-text embedding model with a predictive training objective delivers strong multi-task performance across vision-language benchmarks while keeping model size and architectural changes minimal.
Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers. Trained on a diverse mixture of LatentQA-style tasks (system-prompt QA, binary classification, and a novel self-supervised context prediction), these models accept activation vectors from different layers and generalize to out-of-distribution auditing tasks, outperforming prior white- and black-box methods.
Kling-Omni Technical Report. A unified framework combines a new multimodal vision-language interaction paradigm with a single architecture to perform diverse video generation, editing, and reasoning tasks while improving user control and intention inference.
Seedance 1.5 pro: A Native Audio-Visual Joint Generation Foundation Model. A dual-branch diffusion-transformer with cross-modal integration, post-training optimizations, and an acceleration framework produces tightly synchronized audio-visual outputs.
Adaptation of Agentic AI. A unified taxonomy divides adaptation into agent-focused (A1/A2) and tool-focused (T1/T2) strategies, analyzes trade-offs, and surveys methods, applications, and open challenges for improving agentic AI performance and generalization.
Concerns
People Are Using Sora 2 to Make Disturbing Videos With AI-Generated Kids. Researchers and watchdogs say Sora 2–generated clips depicting photorealistic children in sexualized scenarios are proliferating on social platforms, prompting spikes in reports of AI-created child sexual abuse material and new legal and regulatory responses in the UK and US.
Google’s and OpenAI’s Chatbots Can Strip Women in Photos Down to Bikinis. Users have been leveraging prompts and image-editing features in popular AI chatbots to create nonconsensual bikini deepfakes and are sharing tips to bypass safety guardrails.
The Ultra-Realistic AI Face Swapping Platform Driving Romance Scams. Researchers and payment-tracing firms report the app has generated millions in crypto payments and is being used to create convincing deepfake video chats that facilitate investment and romance fraud across Southeast Asia.
Scammers in China Are Using AI-Generated Images to Get Refunds. Merchants and platforms are seeing a rise in doctored photos and videos—especially for groceries, cheap beauty items, and fragile goods—that buyers use to claim false damage refunds, prompting some sellers to report cases and authorities to detain perpetrators.
Boys at her school shared AI-generated, nude images of her. She was the one expelled. The school initially dismissed students’ reports of AI-generated nude images circulating on social media, investigated without finding the files, and ultimately disciplined the 13-year-old who fought back, while criminal charges were later filed against boys accused of creating and sharing the images.
Policy
New York governor Kathy Hochul signs RAISE Act to regulate AI safety. The law mandates large AI developers disclose safety protocols, report incidents within 72 hours to a new DFS office, and face potential $1M–$3M fines for violations amid industry pushback and calls for federal action.
OpenAI adds new teen safety rules to ChatGPT as lawmakers weigh AI standards for minors. OpenAI updated its Model Spec to impose stricter limits on interactions with users under 18, add real-time safety classifiers and human review for flagged content, and publish AI-literacy resources for teens and parents.
California threatens Tesla with 30-day suspension of sales license for deceptive self-driving claims. A judge found that Tesla used misleading terms like “Autopilot” and “Full Self-Driving” to overstate the capabilities of its driver-assist systems and recommended a 30-day sales-license suspension unless the company makes clearer disclosures and changes its marketing within 90 days.
Analysis
Merriam-Webster names ‘slop’ as its 2025 word of the year. The dictionary’s choice reflected search spikes and public interest in the proliferation of low-quality, AI-produced online content and related cultural trends this year.