🥈 NVIDIA Claims Its GPUs Are a Generation Ahead of Google’s
NVIDIA pushed back after its stock dipped on reports that Meta may explore using Google’s TPUs, insisting its GPUs remain a generation ahead. The company argued its hardware is more flexible and broadly compatible than Google’s ASIC-based TPUs. Google said demand for both platforms is rising. NVIDIA emphasized that scaling laws will continue driving GPU demand across the industry.

💼 MIT Estimates AI Could Replace 12% Of Jobs
MIT’s Iceberg Index finds current AI systems could already perform tasks equal to 11.7% of U.S. jobs. The model maps 151 million workers and 32,000 skills to reveal broad exposure beyond tech roles. Routine functions in HR, logistics, finance and administration show the greatest risk. States like Tennessee and Utah are using the tool to guide training and policy planning.

🤝 Warner Strikes Licensing Deal With Suno
Warner settled its lawsuit against Suno by signing a licensing deal that moves the AI music platform toward paid, rights-cleared song downloads. Artists who opt in will be compensated and keep control of their music and likeness. Suno also acquired Songkick from Warner. The company will phase out its current unlicensed models.

🔄 OpenAI Dial Shift Sparked Safety Overhaul (Paywall)
OpenAI tweaked ChatGPT to boost engagement, rolling out an update that increased daily use but made the model more sycophantic and emotionally validating for some users. That shift coincided with reports of psychological harm and prompted internal studies with MIT. OpenAI rolled back changes, released safer GPT-5, and added detection, break nudges, and personality controls. Lawsuits and debates over growth versus safety continue.

Enjoying our newsletter? Forward it to a colleague—
it’s one of the best ways to support us.

Google’s Gemini 3 narrows the AI gap as a leaked OpenAI memo admits pressure. A breakdown of why Google is now best positioned to win the AI race.

Anthropic detailed a new two-agent harness that helps Claude work reliably across many context windows—an ongoing challenge for long-running tasks that span hours or days. The system pairs an initializer agent, which creates environment scaffolding on first run, with a coding agent that makes incremental, well-documented progress in each subsequent session. The approach countered common failure modes, including attempts to build entire apps in one shot, lost progress between sessions, and prematurely marking features as complete.

Why it matters: long-horizon AI work remains brittle, and this structure helps agents sustain consistent, verifiable progress without relying on persistent memory. Anthropic’s internal tests showed meaningful gains when combining structured feature lists, git-based state tracking, and browser-automation testing. → Read the full article here.

Google’s Gemini 3 debuted with top-tier reasoning performance and was trained entirely on the company’s custom tensor processing units (TPUs), a shift that rattled AI-chip investors. The model’s showing, plus a report that Meta may consider TPUs for 2027 data centers, fueled debate over whether bespoke silicon could chip away at NVIDIA’s dominance. NVIDIA shares fell to a near three-month low, even as the company emphasized its broad software ecosystem and universal model support.

Analysts noted that TPUs may cut costs for hyperscalers building their own large models but remain less attractive for cloud customers who rely on NVIDIA’s CUDA stack. Why it matters: Gemini 3 underscores rising pressure on AI infrastructure economics and signals that custom chips could matter more at the top end—even if GPUs remain essential for most developers and cloud workloads. → Read the full article here.

Google introduced Ironwood, its seventh-generation Tensor Processing Unit, now available to Cloud customers as the company’s most powerful and energy-efficient custom chip to date. Built for high-volume, low-latency inference, Ironwood delivers over 4× the per-chip performance of the prior TPU generation for both training and serving workloads.

The hardware scales to 9,216-chip superpods connected by a 9.6 Tb/s interconnect and sharing 1.77 PB of high-bandwidth memory, reducing training time and energy consumption for large models. Why it matters: Ironwood signals Google’s push to optimize AI infrastructure around inference at scale while tightening integration between model research and chip design. → Read the full article here.

👩‍🦲 AI Scans Your Scalp: Startup MyHair AI uses a custom model trained on 300k images to detect early hair loss and guide treatment, already drawing 1,000 subscribers.

🗣️ ChatGPT Blends Voice and Text: OpenAI’s latest update merges spoken and visual interactions in one flow, a shift the piece says hints at what Apple must match for the next-gen Siri.

🚪 HP Plans Major Job Cuts: HP will eliminate 4,000–6,000 roles by 2028 while leaning on AI to speed development and trim costs, targeting $1B in savings amid rising chip prices.

🔞 Character.AI Shifts Minors to Stories: The company is blocking under-18s from open-ended chats and introducing guided interactive fiction instead, citing safety concerns and new regulations.

❤️ Love Forward Future? Share your unique referral link with friends and colleagues to unlock exclusive Forward Future perks! 👉 Get your link here.
📢 Want to advertise with Forward Future? Reach 600K+ AI enthusiasts, tech leaders, and decision-makers. Let’s talk—just reply to this email.
📥 Got a hot tip or burning question? Drop us a note! The best reader insights, questions, and scoops may be featured in future editions. Submit here.
🛰️ Want more Forward Future? Follow us on X for quick updates, subscribe to our YouTube for deep dives, or add us to your RSS Feed for seamless reading.

Thanks for reading today’s newsletter—see you next time!
Matthew Berman & The Forward Future Team

🧑‍🚀 🧑‍🚀 🧑‍🚀 🧑‍🚀