Good morning. It’s Friday, September 26th.

On this day in tech history: In 2013, Facebook announced its DeepFace system, achieving 97.35 percent accuracy on the Labeled Faces in the Wild dataset, nearly matching human-level facial recognition performance at the time. The system used a nine-layer deep neural network with 120 million parameters and a 3D alignment step to normalize facial poses before feature extraction.

In today’s email:

• ChatGPT debuts ‘Pulse’

• DeepMind’s Gemini 1.5 lets robots plan, reason, and search the web for tasks

• Meta arms AI with agent entropy handling and execution-tracking LLMs

• Copilot expands with Claude and GPT orchestration while Microsoft tests AI content marketplace

• 5 New AI Tools

• Latest AI Research Papers

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

GPT-5 matches human pros in 40% of real-world tasks, ChatGPT debuts ‘Pulse’

GPT-5-high now rivals human experts in 40.6% of tasks on OpenAI’s new GDPval benchmark, covering 44 roles across nine industries from healthcare to finance and manufacturing. Up from GPT-4o’s 13.7%, the gains reflect sharper applied reasoning, faster research synthesis, and more precise report generation. Claude Opus 4.1 hits 49%, boosted by polished visuals, showing how presentation can sway benchmark outcomes.

For users, ChatGPT Pulse upgrades your assistant from reactive to proactive, scanning your chat history, emails, and apps like Google Calendar overnight to deliver personalized visual “cards” each morning. From travel plans and workouts to meals, meetings, and task tracking, Pulse anticipates your day and adapts to your context. With feedback and the “Curate” tool, you fine-tune what it surfaces, giving a real-world glimpse of autonomous AI that acts intelligently without being prompted.

Monetization is on the radar, too. The company is hunting for an ad lead to steer subscriptions, ad products, and other revenue streams under Fidji Simo, who now runs everything outside research, infrastructure, hardware, and security.

Meanwhile, ChatGPT users have spotted “alpha” agent variants, “truncation” and “prompt expansion,” that auto-activate browsing and tools. Early signs point to GPT-5 agents moving from reactive chatbots to proactive, autonomous task managers. Read more.

DeepMind’s Gemini 1.5 lets robots plan, reason, and search the web for tasks

Google DeepMind’s latest Gemini Robotics 1.5 and Robotics-ER 1.5 models are taking robots beyond single-step instructions, letting them plan multistep actions, adapt to real-world complexity, and pull guidance from the web. ER 1.5 interprets environments and converts web insights into stepwise instructions, which Robotics 1.5 executes via advanced vision-language processing. Skills transfer across platforms, so dual-arm ALOHA2 routines work on Apollo’s humanoid frame.

Gemini 2.5 Flash and Flash-Lite sharpen agentic reasoning for software workflows, cutting output tokens by 24–50% to reduce latency and compute costs. Flash gains 5% on SWE-Bench Verified, and both models now follow complex instructions with higher fidelity, support richer multimodal outputs, and handle extended task chains, letting autonomous agents execute intricate sequences with minimal human oversight.

NotebookLM is also evolving. Users can soon pull sources directly from Google Drive and, via a “deep research” mode, query the web for multi-source references. Hints of audio and video overview generation suggest teams may soon produce multimedia briefings straight from their notes.

Productivity tools are catching up too. Google Vids converts Slides decks into video with autogenerated scripts, narration, and background music, while Veo 3 adds AI-generated clip storyboarding with human oversight.

Search Live, embedded in Lens, turns queries into interactive, task-oriented AI guidance.

Also, a new 2025 DORA Report finds 90% of coders use AI daily, but only 24% trust it; Gemini Code Assist generates two hours of work per engineer per day, powering over a quarter of Google’s new code. The report recommends “small-batch workflows” and guardrails to balance speed with reliability. Read more.

Meta arms AI with agent entropy handling and execution-tracking LLMs 

Meta is sharpening AI for real-world complexity with three major moves. The Agents Research Environment (ARE) and Gaia2 benchmark stress-test agents in dynamic simulations packed with stateful apps, async events, deadlines, API failures, and multi-agent coordination. Gaia2 scores resilience and accuracy, not just output correctness. Early runs across 1,120 mobile tasks put GPT-5 at the top, setting a high bar for multi-agent robustness.

On coding, Meta’s 32B-parameter Code World Model (CWM) goes beyond code generation. Trained on 120M Python execution traces and fine-tuned with reinforcement learning across coding, math, and reasoning tasks, it simulates execution line by line, predicts infinite loops, runtime behavior, and algorithm complexity. Benchmarks show 65.8% pass@1 on SWE-bench Verified, 68.6% on LiveCodeBench, 96.6% on Math-500, and 76% on AIME 2024, with a 131k-token context window and open checkpoints for experimentation.

Meta is also launching Vibes, a new feed in its Meta AI app and on meta.ai that generates short-form AI videos for users to browse, remix, and share.

The feature draws on AI models from Meta as well as partnerships with Midjourney and Black Forest Labs, producing content ranging from surreal animations to historical vignettes with minimal human input. Read more.

Copilot expands with Claude and GPT orchestration while Microsoft tests AI content marketplace

Copilot is no longer bound to a single engine. Anthropic’s Claude Sonnet 4 and Opus 4.1 are now embedded alongside OpenAI’s GPT models across Microsoft 365 Copilot and Copilot Studio. In practice, that means a research agent drafting a market report can run on Claude for sharper reasoning, or on GPT for faster narrative polish.

Copilot Studio takes it further with orchestration, letting enterprises assign different models to different steps of a workflow. A compliance-heavy check might sit on Claude, customer-facing copy on GPT-5, all stitched together into a multi-agent system with fallback logic if one model is disabled. Early rollout is already live, with preview in two weeks and production by year’s end.

On the content side, Microsoft is piloting a Publisher Content Marketplace (PCM), pitched to outlets in Monaco last week. Instead of flat fees, PCM would compensate publishers based on content usage inside AI products like Copilot, a step toward usage-linked royalties in an industry still dominated by one-off licensing deals. Read more.

• Moonshot AI’s Kimi assistant, ‘OK Computer’ offers agent mode for creating multi-page websites, slides

• Intel seeks investment from Apple

• Alibabas Wan2.5-Preview lets users turn photos and text prompts into videos with matching audio

• AI can now pass the hardest level of the CFA exam in a matter of minutes

• Perplexity released the Search API

• Qualcomm announces Snapdragon 8 Elite Gen 5 mobile chipset

• Cloudflare releases NET Dollar stablecoin secure transactions for the agentic web

• Boston Dynamic has developed a robotic gripper arm for logistics

• Spotify is finally taking steps to address its AI slop and clone problem 

• Ardent AI beats the odds to launch world's 'first agentic engineer' for data pipeline maintenance

• Musk’s Grok cleared for use across US government agencies

• Adobe adds Nano Banana to Photoshop Beta and expands Firefly with Runway Aleph and Moonvalley’s Marey models

• Even Disney reportedly lacks enough data to train a top-tier AI video model

• Open source Qwen3-VL outperforms Gemini 2.5 Pro in major vision benchmarks, Alibaba reports

• Silicon Valley’s latest anti-regulation argument: AI laws will summon the antichrist

• Alibaba inks Nvidia deal, expands data centers, unveils most powerful AI model yet

• Tencent’s new AI technique teaches language models ‘parallel thinking’

• Apple’s memory-saving AI breakthrough could save enterprises millions

• Runway launches API Playground

• Meet the omni-bodied robot that learns in real time and keeps the robot moving even when its limbs are destroyed

• A startup used AI to make a psychedelic without the trip

• Hugging Face adds LIBERO benchmark, unlocking 130+ vision-language-action tasks

• Krea AI adds Kling 2.5, enhancing hyper-realistic motion capabilities

• Higgsfield AI launches unlimited WAN 2.5 with sound and free daily generations

• LexisNexis exec says it’s ‘a matter of time’ before attorneys lose their licenses over using open-source AI pilots in court

• Mistral AI tests Tone controls and custom styles in Le Chat

• Neon, the No. 2 social app on the Apple App Store, pays users to record their phone calls and sells data to AI firms

• 80% of Gen Z and millennials are turning to AI for financial advice, but more than half say they've made a poor decision or mistake as a result

• US firm's humanoid robots to transform shipyard operations

• Brex turns accounting into a one-click setup with Puzzle integration for startups

• FICO’s answer to AI risk: A foundation model that scores every output for accuracy and compliance

📄 Video models are zero-shot learners and reasoners

📄 CAR-Flow: Condition-Aware Reparameterization Aligns Source and Target for Better Flow Matching ​​

📄 Soft Tokens, Hard Truths

📄 SIM-CoT: Supervised Implicit Chain-of-Thought 

📄 Reinforcement Learning on Pre-Training Data

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!