Good morning. It’s Friday, August 29th.

On this day in tech history: In 1993, the Mbone (“multicast backbone”) quietly proved the Internet could move more than text. It streamed a David Blair’s full-length film, Wax or the Discovery of Television Among the Bees to thousands using IP multicast, tunneling packets through routers like a guerrilla overlay network. The demo was fragile, but its elegant overlay proved large-scale media wasn’t a server problem - it was a distribution problem.

In today’s email:

• GPT-Realtime

• Google’s Multimedia AI

• Musk’s Big Week

• 5 New AI Tools

• Latest AI Research Papers

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

OpenAI rolls out gpt-realtime and GPT-5 Codex but joint safety tests expose cracks

The new gpt-realtime model is now production-ready through the Realtime API. Unlike older systems, it processes speech directly without text conversion, cutting latency and capturing nuance like laughter, sighs, or accent shifts. It can even switch languages mid-sentence.

Benchmarks show major jumps: 82.8% on Big Bench Audio (vs. 65.6% prior) and 30.5% on MultiChallenge (vs. 20.6%). Developers also get SIP support for phone systems, MCP for secure tool access, image input for contextual analysis, and two new expressive voices (Cedar, Marin). Pricing dropped 20% to $32 per million input tokens.

On the dev side, ChatGPT Codex now runs on GPT-5, integrated directly with GitHub for pull requests, branch management, and code reviews. A new IDE extension, upgraded CLI, one-shot task execution, and customizable “agents.md” files streamline automation.

But joint OpenAI–Anthropic tests using the SHADE-Arena sabotage framework flagged risks: GPT-4.1 and GPT-4o still leaked detailed misuse instructions under pressure, while Claude models leaned toward refusals but showed sabotage quirks. Both families exhibited sycophancy, validating unsafe user decisions. With GPT-5 evaluations ahead, audits and refusal testing remain non-optional. Read more.

From AI avatars to private blockchain, Google layers multimedia and fintech tools

Google Vids now supports AI avatars that generate videos from scripts with selectable voices and personas, automatic transcript trimming to remove filler words, and eight-second image-to-video clips powered by Veo 3. Complementing this, Flow offers five (5) free Veo 3 Fast AI videos, or a single standard video, per month, with credits and per-second API pricing ($0.040 Fast, $0.0075 standard) for scalable production.

On the learning front, NotebookLM is testing Drive search and AI Flashcards. “Discover Sources” surfaces internal Docs and Slides alongside web content, while AI Flashcards generate study aids from documents, extracting key facts and questions. Public sharing of audio artifacts and upcoming video overviews deepen collaboration and multimedia integration.

Meanwhile, Google Cloud is quietly piloting its own blockchain, Google Cloud Universal Ledger (GCUL), a private, permissioned Layer 1 network for Python-based smart contracts, payment automation, and digital asset management. Initially tested with CME Group, GCUL offers a “credibly neutral” infrastructure for banks and fintechs, emphasizing compliance, API accessibility, and regulated adoption, though its private design has sparked debate over decentralization. Read more.

Musk’s week: a coding agent, a vision-trained robot, and a reusable Starship

Elon Musk’s empire just pulled a triple tech swing that feel like pressure tests for scale. At xAI, the new grok-code-fast-1 model dropped, a lean, agentic coder built to churn out routine dev tasks quickly and cheaply. Unlike bloated LLMs, it’s tuned for compact execution, already plugged into GitHub Copilot and Windsurf. Free to try (for now), the real play is whether it can evolve from quick snippets into full-stack automation that seriously rivals Microsoft’s Copilot and OpenAI’s Codex.

Tesla’s Optimus, meanwhile, just ditched motion-capture suits for helmet-and-backpack rigs with five synced cameras, training Optimus humanoids via vision-only feeds. It’s the Autopilot playbook: flood models with scaled human video instead of costly teleoperation. Gains include: data density, speed, lower cost. Risks: zero tactile feedback, making dexterity harder. Leadership shuffled too, with AI chief Ashok Elluswamy folding Optimus under Tesla’s camera-first AI stack.

And then there’s SpaceX. Starship S37 pulled a precision splashdown after a 66-minute hop, validating reinforced heat tiles, engine-out recovery, and a satellite bay. Both stages survived, underscoring that fully reusable rocketry is shifting from theory to engineering cadence. Read more.

• Anthropic’s settlement with authors may be the ‘first domino to fall’ in AI copyright battles

• Meta races to patch Llama AI while Hypernova glasses stay in test mode

• Anthropic users face a new choice – opt out or share your data for AI training

• Google and Grok are catching up to ChatGPT, says a16z's latest AI report

• Microsoft drops first in-house AI models, built for speed not show

• Nous Research drops Hermes 4 AI models that outperform ChatGPT without content restrictions

• MathGPT.ai, the ‘cheat-proof’ tutor and teaching assistant, expands to over 50 institutions

• AI opens door for new finance skillsets 

• Scientists develop interface that ‘reads’ thoughts from speech-impaired patients

• Watch: World's first flying car built by US firm to start operations at Silicon Valley airports

• Over half of professionals think AI trainings feel like a second job, LinkedIn survey finds

• PixVerse debuts AI video model V5 with free access week, surpasses 100M global users

• Watch: China open-sources HunyuanVideo-Foley for high-fidelity TV2A audio

• Chip giant Nvidia beats revenue expectations, defying fears of AI 'bubble'

• US fighter pilots try taking directions from AI for the first time

• Packaging with functional inks and AI recognition shows product condition in real time

• Salesforce builds 'flight simulator' for AI agents as 95% of enterprise pilots fail to reach production

📄 Understanding Tool-Integrated Reasoning

📄 Self-Rewarding Vision-Language Model via Reasoning Decomposition ​​

📄 Mixture of Contexts for Long Video Generation

📄 rStar2-Agent: Agentic Reasoning Technical Report

📄 MIDAS: Multimodal Interactive Digital-human Synthesis via Real-time Autoregressive Video Generation

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!