AI News · November 20, 2026 · 7:28

Google launches Gemini 4 Neon & Gemini Carbon slowness and Antigravity cache - AI News (Nov 20, 2026)

An AI-invented CPU, Gemini 4 Neon, OpenAI's $23K-a-day research agents, Anthropic's compute surge, and the debate over handing control to AI.

Google launches Gemini 4 Neon & Gemini Carbon slowness and Antigravity cache - AI News (Nov 20, 2026)
0:007:28

15 free episodes a day, 5 editions, 3 languages, about $4 a day to produce. A coffee keeps it going ☕

Support the show →

Our Sponsors

Today's AI News Topics

  1. Google launches Gemini 4 Neon

    — Google released Gemini 4 Neon, a fast frontier model for coding, knowledge work, science and cybersecurity, with a 1 million-token context window. It is available in the Gemini API, AI Studio and Google Antigravity.
  2. Gemini Carbon slowness and Antigravity cache

    — Kernel found Gemini 4 Carbon's slow bug fixing comes mostly from overthinking and inefficient tool use. A separate analysis traced Antigravity quota drain to a prompt cache that expires after about five minutes.
  3. Stitch Loop and Genie 4 updates

    — Google Labs introduced Stitch Loop, an agent that keeps improving shipped software through pull requests. Google also made Genie 4 available to AI Ultra subscribers for prompt-driven interactive worlds with adaptive audio.
  4. OpenAI agents accelerate internal research

    — OpenAI reports that coding agents now complete 46% of two-to-four-hour research tasks, with heavy users spending tens of thousands of dollars a day. Safety and alignment work nearly doubled, and human oversight remains essential.
  5. AI designs chips, including Frostbite CPU

    — OpenAI models improved power, performance and area in its Jalapeño inference chip and in AMD designs, with Synopsys EDA integration planned. Architect Labs says AI designed and verified the Frostbite CPU architecture end to end, claiming a 4.59x performance-per-watt gain.
  6. Meta agents spread ads ranking techniques

    — Meta uses AI agents for more than 60% of technique propagation across its ads ranking models. The company reports 3x engineering productivity and more than $45 million in estimated compute savings.
  7. Anthropic and OpenAI compute spending surge

    — Epoch AI estimates that Anthropic's compute spending reached nearly $10 billion in Q3 2026, growing about 4x a year. That outpaces OpenAI's roughly 3x annual growth, showing how capital-intensive the frontier AI race has become.
  8. Inside the Codex sandbox

    — OpenAI explains how the Codex sandbox restricts filesystem and network access through platform-native sandboxing on macOS, Linux and Windows. The approach adds defense in depth to agentic coding.
  9. Local AI and Chrome bandwidth concerns

    — Pete Warden argues that AI increasingly wants to run locally, citing cost, privacy and fast-improving open models. Meanwhile, a user traced a Chrome bandwidth spike to on-device AI features that are enabled by default.
  10. Safeguards blocking defensive security work

    — A developer says Anthropic's Claude Code safeguards blocked a legitimate security fix. They switched to EU-hosted open-weight models, reigniting debate over AI safety and defensive cybersecurity.
  11. AI gig work for experts

    — Platforms like Scale AI's Outlier hire doctors, lawyers and PhDs to train AI, with pay that varies sharply by country. Critics call this AI-era gig work a new form of labor extractivism.
  12. Silent consensus and shorter AI timelines

    — Dean W. Ball warns of a 'silent consensus' in which humans gradually cede control to AI. Zvi Mowshowitz notes that timelines have shortened since the METR report, with real odds of automated AI research by 2027.

Sources & AI News References

Full Episode Transcript: Google launches Gemini 4 Neon & Gemini Carbon slowness and Antigravity cache

An AI system has reportedly invented an entirely new CPU architecture on its own, from the instruction set all the way down to the final chip layout. It also claims more than four times the efficiency of a modern core. We'll get to that. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is November 20th, 2026. On today's list: a new Gemini model, OpenAI's report on how agents are speeding up its own research, a big jump in compute spending, and a sharp essay on whether humans should keep the wheel. Let's get going.

Google launches Gemini 4 Neon

We'll start with Google, which has released Gemini 4 Neon. The new model is built to finish complex software engineering, knowledge work and science tasks quickly. Google points to strong benchmark results across coding, legal, tax and scientific reasoning, a 1 million-token context window, and less fabrication. Cybersecurity gets particular attention. Google says Neon finds, validates and patches vulnerabilities at lower cost than rival frontier models. It's rolling out now in the Gemini API, AI Studio and Google's business and consumer products.

Gemini Carbon slowness and Antigravity cache

Two updates on Google stories we've been following. Kernel dug into why Gemini 4 Carbon spends close to an hour on bug fixes that other models finish in minutes. Slower inference is only part of the answer. The bigger problem is behavior. Carbon audits unrelated code, guards against edge cases it has already ruled out, and re-reads files again and again. In one run it found the bug early, then waited a long time before making its first edit. On the Antigravity quota drain, new measurements put the prompt cache expiry between four minutes forty-five seconds and five minutes. The same analysis found that Google already built a keep-alive system but appears to have switched it off for Gemini models. A well-timed ping can save a lot of tokens, but pinging through long idle stretches just wastes them.

Stitch Loop and Genie 4 updates

Google Labs also introduced Stitch Loop, a long-running agent that connects to a codebase and live product metrics. Its job is to keep improving software after launch, on things like latency, accessibility or checkout conversion. It writes its changes in a sandbox and opens pull requests for humans to review. On the creative side, Genie 4 is now available to AI Ultra subscribers. Users can reshape interactive worlds in real time with plain-language prompts, and the worlds now come with adaptive audio and a community gallery for sharing and remixing.

OpenAI agents accelerate internal research

Over to OpenAI, which published a candid look at how agents are changing its own research. Agents now complete nearly half of two-to-four-hour research tasks on their own, up from about a quarter in March. They're also expensive. The median researcher uses roughly $2,400 a day at API list prices, and heavy users reach about $23,000. OpenAI says agent spending now runs at about twice researchers' non-equity pay. Safety and alignment work nearly doubled over the year, especially after the July Hugging Face incident. The takeaway is real acceleration that is uneven across tasks and still depends on human oversight.

AI designs chips, including Frostbite CPU

OpenAI also says its models are now contributing to chip design. Model-generated changes improved power, performance and area in its custom Jalapeño inference chip and in AMD design files, and some of those gains are already proven in silicon. OpenAI is partnering with Synopsys to bring these models into standard chip design tools. That brings us back to the opening tease. Architect Labs says it built Frostbite, a general-purpose CPU architecture discovered, designed and verified end to end by AI. The company reports a 4.59x gain in performance per watt over a modern in-order core. If that holds up under outside scrutiny, it marks a big step from AI as a coding assistant to AI as an architect.

Meta agents spread ads ranking techniques

Meta, meanwhile, is using AI agents to carry proven improvements from one ads ranking model to the rest of its fleet. That job used to be slow, manual work. Agents have handled more than 60 percent of it since March, and Meta reports triple the engineering productivity and over $45 million in estimated compute savings. The goal is to cut rollout time from more than a year to about three months.

Anthropic and OpenAI compute spending surge

All of this costs money. Epoch AI estimates that Anthropic's compute spending reached nearly $10 billion in the third quarter. That's up from $433 million in 2023, which works out to roughly fourfold growth each year. OpenAI is growing at about three times per year, though Epoch notes its data on OpenAI is thinner.

Inside the Codex sandbox

On safety in practice, OpenAI explained how the Codex sandbox works. By default, Codex can write only inside your workspace and has no network access. It uses the security features built into each operating system rather than containers, so developers can keep working in their usual setup. OpenAI admits some risk remains, but the layered approach shows how agentic tools are being made fit for real development work.

Local AI and Chrome bandwidth concerns

Pete Warden makes the case that AI wants to be local. Open models are often only a few months behind commercial systems, and running them on your own hardware can be cheaper, more private and less restrictive. Local doesn't always go smoothly, though. One user traced a sudden spike in Chrome bandwidth to on-device AI features that were switched on by default, and is now asking why browsers ship them that way.

Safeguards blocking defensive security work

A related complaint comes from a developer whose attempt at a simple security fix was blocked by Claude Code's safeguards. They describe the exemption process as slow and invasive, and say they ended up using EU-hosted open-weight models to get the fix done. Their argument is that restrictions hold back defenders while attackers simply use other tools.

AI gig work for experts

On labor, a new essay looks at platforms like Scale AI's Outlier, which hire doctors, lawyers and PhDs to rank and correct AI outputs. Pay varies widely, from high advertised rates in the U.S. to far less in the Philippines or Tanzania, and workers can lose access to tasks without warning. The author argues that AI is breaking skilled professions into gig work.

Silent consensus and shorter AI timelines

Finally, the bigger picture. Dean W. Ball warns against what he calls the 'silent consensus,' the assumption that humans will gradually hand control to AI simply because it's more efficient. He wants human accountability over AI agents, limited permissions, and tools that strengthen people rather than replace them. Zvi Mowshowitz, meanwhile, says timelines have shortened since the METR report. He sees a substantial chance, though still below even odds, of AI fully automating AI research by 2027, and he argues that plans assuming several more years are now risky.

That's the rundown for November 20th, 2026. Links to all stories can be found in the episode notes. I'm TrendTeller. Thanks for listening to The Automated Daily, AI News edition, and see you tomorrow.

More from AI News