Good morning. It’s Friday, October 10th.

On this day in tech history: In 1997, Hochreiter & Schmidhuber introduced Long Short-Term Memory, a gated RNN architecture that overcame vanishing gradients by preserving information through memory cells. This design enabled learning over long time spans, revolutionizing sequence modeling in speech, handwriting, and NLP, and paving the way for GRUs, peepholes, and attention-based models.

In today’s email:

• Figure 03 Robot

• OpenAI’s Sora Success

• Google’s Stitch

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

Meet the next wave of humanoids. Smart, strong, and ready for anything

Humanoid robots are leveling up fast, and three new models released this week show just how far we’ve come.

Figure AI’s Figure 03 stands 5’6” and pairs Helix AI with dual palm cameras, fingertip sensors sensitive to just three grams, and a 60% wider visual field. It folds laundry, sorts parts, and handles fragile items with near-human precision. Actuators hit human pick-and-place speeds, wireless foot charging keeps it running, and upgraded speakers and mics make conversations feel natural. Built for scale with die-casting and injection molding at BotQ, Figure 03 targets 12,000 units a year, though consumer deployment remains pending.

Boston Dynamics’ Atlas just got a dexterity upgrade. Its three-fingered gripper has seven degrees of freedom, an articulated thumb, tactile fingertips, and palm cameras. Paired with Toyota’s Large Behavior Model, Atlas can sort, lift, and organize objects while adapting its grip on the fly. Mobility plus fine manipulation makes it a serious contender for industrial and domestic work, even if it’s still slower than humans.

China’s DR02 brings resilience to extreme environments. IP66-certified, it runs from -20°C to 55°C, lifts 10 kg per arm (20 kg total), and climbs stairs or 20° slopes at 1.5 m/s. A 275 TOPS processor and modular limbs mean it’s reliable, low-maintenance, and ready for hazardous industrial sites.

From folding laundry to braving extreme outdoor conditions, these humanoid robots are no longer just ideas. They’re moving, sensing, and adapting in the real world. Read more.

OpenAI’s Big Week brings 1M Sora downloads, GPT-5 Pro milestone, locks 26GW compute

OpenAI’s invite-only AI video app Sora has stormed Apple’s App Store, hitting over 1 million downloads in under five days, outpacing ChatGPT’s 2023 debut. The iOS-exclusive app generates short videos from text prompts, though rapid adoption has triggered copyright complaints from media groups, prompting OpenAI to give rights holders more control and refine content policies.

Behind the consumer buzz, OpenAI is massively scaling infrastructure through its $500 billion Stargate program, securing over 26 gigawatts of AI compute via partnerships with Nvidia, AMD, Oracle, and SoftBank. AMD is offering OpenAI up to 10% equity for co-designing next-gen GPUs, while Nvidia’s $100 billion investment effectively makes it a stakeholder in OpenAI. Sam Altman confirmed more deals are on the horizon, noting next-gen models could demand $60 billion per gigawatt of compute. OpenAI also expanded its ChatGPT Go subscription plan to sixteen Asian markets, offering higher message and image limits and localized pricing.

Meanwhile, GPT-5 Pro set new ARC-AGI records. 70.2% on ARC-AGI-1 and 18.3% on ARC-AGI-2, averaging $4.78 per task. Despite this momentum, Google’s TPU infrastructure and early GPU investments give it a cost and scale advantage. OpenAI’s reliance on third-party infrastructure raises long-term sustainability questions, even as its products push the envelope on adoption, consumer engagement, and AI capability. And the legal stakes rise. ChatGPT logs and AI-generated art are now evidence in California’s 2025 Palisades Fire case, underscoring AI’s impact on real-world accountability. Read more.

Google Stitch brings low-code prototyping, design control, and memory to AI tools

Google is doubling down on AI across design, development, and reasoning. Its design platform Stitch is testing three new modes: Annotate, which applies contextual UI edits via the Gemini-powered Nano-Banana model; Theme, giving full control over colors, typography, and components; and Interactive, enabling low-code UX prototyping with click, input, and describe prompts. These upgrades push Stitch toward an AI-native Figma alternative, especially for teams already in Google Cloud.

Prompted Variants takes Google’s design AI up a notch. Designers can now tweak UIs with prompts like “make checkout easier” or “show me different layouts,” or skip the prompt and let Stitch rethink the screen entirely. In early demos, full makeovers happen in minutes.

Developers get smarter workflows with Gemini CLI extensions. They now integrate AI reasoning and automation with popular tools like Figma, Stripe, Shopify, Postman, and Snyk. Partner launch extensions add observability, CI/CD automation, and security scanning. Prebuilt playbooks teach the AI how to use APIs, letting over a million developers build context-aware, self-updating toolchains.

Gemini 2.5 Deep Think also set new benchmarks in FrontierMath, excelling in geometric reasoning while still hitting limits in numeric precision. Google and the University of Illinois also introduced ReasoningBank, giving LLMs structured “memories” of past reasoning. Paired with Memory-aware Test-Time Scaling, agents can explore multiple reasoning paths, boosting performance up to 8.3 points while cutting unnecessary compute.

Consumer AI is expanding too. Virtual try-on now includes shoes in Australia, Canada, and Japan, and AI-powered Search Mode is live in 40+ European countries, delivering context-aware, conversational results. From design to coding to reasoning, Google is shaping a fully integrated, self-improving AI ecosystem. Read more.

• Nvidia backs xAI with $2B as funding hits $20B, calls AMD-OpenAI deal “clever”

• Anthropic unveils modular plugin system for Claude Code

• Google and AWS Bring AI Agents Into the Office

• One decoder to rule them all. Meta AI’s OpenZL makes compression modular

• China markets reopen to show if AI buzz outshines soft spending

• When chatbots speak, who’s liable? The law isn’t sure anymore 

• MIT doubles optical atomic clock precision with quantum trick

• A whopping 72% of S&P 500 companies disclosed AI as a 'material risk' on their 10-Ks this year

• This startup wants to spark a US DeepSeek moment

• Billionaire investor Marc Andreessen says AI destroying jobs and making everyone poor is a 'fallacy'—and even if that did happen, prices would drop

• 1Password says it can fix login security for AI browser agents

• SoftBank makes a $5.4 billion bet on AI robots

• Without data centers, GDP growth was 0.1% in the first half of 2025, Harvard economist says

• Olympics-Google to power LA28 with cloud infrastructure, AI tools 

• Zendesk says its new AI agent can solve 80% of support issues

• Hardware vulnerability allows attackers to hack AI training data

• New AI tool opens 3D modeling to blind and low-vision programmers

• China proposes global drive to build AI-powered satellite mega network for all

• Intel unveils new processor powered by its 18A semiconductor tech

• MIT: Using generative AI to diversify virtual training grounds for robots

• OneDrive is getting a new Windows app and an AI photo agent

• Billion-dollar tech company Starkey unveils a hearing aid with AI accessible enough for all generations to use: 'The ear is the new wrist'

• These founders argue their tiny AI models are better for customers and the planet

• Internal documents show Oracle's average profit margin for AI cloud deals was 16%

• The Browser Company’s Dia is now available on Mac, no invite needed

• Figma partners with Google to add Gemini AI to its design platform

• AI tools can help hackers plant hidden flaws in computer chips, study finds

• Deloitte’s AI slip-up reveals bigger threat to consulting firms

• Echelon's AI agents take aim at Accenture and Deloitte consulting models

• What MIT got wrong about AI agents: New G2 data shows they’re already driving enterprise ROI

📄 Fast-dLLM v2: Efficient Block-Diffusion LLM

📄 CoDA: Coding LM via Diffusion Adaptation  ​​

📄 Less is More: Recursive Reasoning with Tiny Networks  

📄 Scientific Algorithm Discovery by Augmenting AlphaEvolve with Deep Research 

📄 In-the-Flow Agentic System Optimization for Effective Planning and Tool Use

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!