Audio King of AI? Figure Robot PWNS Critic; Bye Bye CFAs

Today's AI forecast: 🌥️

The Future Is Listening

AI’s next interface shift may not be visual but audible.

On the No Priors podcast, ElevenLabs CEO Mati Staniszewski laid out the case that voice is becoming the default interaction layer for AI as systems move into real time and multimodal territory.

ElevenLabs is already doing about $300M ARR, with roughly 5M monthly active users across creators and enterprises. The company pairs research-grade voice models with production-ready tools, but the bigger story is that the voice market itself is expanding faster than any one player.

This is not about talking chatbots ...voice is slipping into infrastructure. Support centers are morphing into always-on conversational agents. Education is leaning toward persistent AI tutors. Commerce is experimenting with voice-led discovery and conversion. As the stack matures, raw model quality is getting commoditized, and competition is shifting elsewhere.

Why it matters

The real moat in voice AI is no longer realism. It’s latency, orchestration, voice libraries, compliance and distribution. Big Tech and open source are moving aggressively into the space, compressing advantages at the model layer. The companies that win will be the ones that make voice boringly reliable, legally safe, and everywhere.

The Deets

• ElevenLabs leads today on quality and polish, not exclusivity
• Market pressure is pushing voice from novelty to invisible infrastructure
• Competitive battles are moving downstream into systems, not sounds

Key takeaway

Listening is the next interface war, and it’s going to get loud.

🧩 Jargon Buster: Multimodal - AI systems that can process and generate multiple input types like text, audio, images, and video at once.

Source: AI Secret

⚡ Power Plays

NVIDIA Moves In on the Agent Stack

NVIDIA just open-sourced Nemotron 3, a model family built specifically for agentic AI, not general chat. The lineup spans Nano (30B), Super (100B), and Ultra (500B), with Nano available now and larger models coming in 2026. Unlike most U.S. competitors, NVIDIA is publishing training data, fine-tuning tools, RL environments, and eval suites alongside the models.

Nemotron uses a hybrid latent mixture-of-experts architecture, optimized for multi-agent coordination, long-horizon reasoning, and low-cost inference. Early adopters include Cursor, Perplexity, ServiceNow, and CrowdStrike.

Why it matters

This is NVIDIA quietly becoming the default infrastructure layer for agents, not just the GPUs they run on. Nemotron lets teams spin up dozens or hundreds of agents without lighting frontier-model money on fire. Enterprises get inspectable, sovereignty-friendly models. Startups get a practical backbone that only calls proprietary models when needed.

The Deets

• Nano outperforms similarly sized open models like Qwen3-30B on coding
• Responses generate 3x faster at comparable sizes
• Fully open release keeps Western developers competitive while tying them to NVIDIA hardware

Key takeaway

NVIDIA is not chasing chatbot hype, but swallowing the agent stack from the bottom up.

🧩 Jargon Buster: Agentic AI - AI systems designed to plan, act, and coordinate across tools and environments with minimal human input.

Source: AI Secret, The Rundown AI

🤖 Robotics

Figure Challenged... Responds: 'Hold My Beer'

The internet asked whether Figure’s humanoid robot could survive real-world chaos, not a polished demo. AI commentator Mehmet Aykul challenged the company to show its robot handling “10,000 messy packages without failing.” Figure founder Brett Adcock responded with receipts: a full 60-minute, uncut video of the robot sorting bags and boxes nonstop. No edits. No music. Every label faced down. Every motion deliberate.

This wasn’t a one-off flex. It follows an 11-month pilot with BMW, where Figure 02 loaded more than 90,000 parts for X3 assembly lines. The video effectively closes the debate on whether the robot can do the work.

Why it matters

The conversation around humanoids just shifted. Motion, dexterity, and reliability are no longer the bottleneck for Figure. The new question is manufacturing. Can Figure turn proof into production before patience runs out? Investors and factories are done watching demos. They want delivery schedules.

The Deets

• 60 minutes of uninterrupted real-world operation
• Proven performance in automotive manufacturing
• Credibility risk replaced by execution risk

Key takeaway

Everyone believes Figure can work. Now it has to clock in at scale.

🧩 Jargon Buster: Pilot Program - A limited deployment used to validate performance before full-scale rollout.

Source: Robotics Herald

iRobot Finally Runs Out of Floor to Clean

After 35 years as the face of consumer robotics, iRobot, maker of Roomba, has filed for Chapter 11 bankruptcy. The company will be acquired by its contract manufacturer, China’s Picea Robotics. iRobot says operations, apps, and support will continue for now.

Once a category-defining company, iRobot was squeezed by cheaper Chinese competitors, U.S. tariffs, and the collapse of its blocked Amazon acquisition, which removed its last realistic escape hatch.

Why it matters

This is what happens when pioneers stop outpacing imitators. While Roborock and Ecovacs built vertically integrated stacks with AI mapping and aggressive hardware iteration, iRobot spent years defending margins in a commoditized market. By the time the robots got smarter, the business model didn’t.

The Deets

• Brand survives, independence does not
• Acquired by its own manufacturer
• Consumer robotics proves brutally unforgiving

Key takeaway

The robot that taught the world to trust home automation couldn’t adapt fast enough to survive it.

🧩 Jargon Buster: Chapter 11 - A form of bankruptcy that allows a company to restructure while continuing operations.

Source: Robotics Herald

🏢 Power Plays

IBM Drives Dull Into Enterprise

IBM open-sourced CUGA Agent, a Configurable Generalist Agent designed for enterprise workflows, not demos. Task completion rates sit around 60% on web benchmarks and 50% on API-heavy tasks. Numbers are fairly strong under real-world constraints.

Instead of chasing intelligence, IBM focused on explicit planning, policy compliance, sandboxed execution, and re-planning when things fail.

Why it matters

Enterprises care less about brilliance than not breaking production systems. CUGA makes failure observable, contained, and reversible, which gives IT teams something they can actually deploy. In regulated environments, trust is the real bottleneck.

The Deets

• Emphasis on control over autonomy
• Designed to fail safely, not magically succeed
• A sharp contrast to black-box, fully autonomous agent bets

Key takeaway
In enterprise automation, survivability beats cleverness every time.

🧩 Jargon Buster: Sandboxed Execution
Running AI actions in isolated environments so failures cannot impact live systems.

Source: AI Secret

🧪 Research & Models

AI Reasoning Models Smash CFA Exams

A new study found that six leading AI models now pass all three levels of the CFA exams. Gemini 3.0 Pro posted a record 97.6% on Level I, while GPT-5 topped Level II at 94.3%. Just two years ago, GPT-3.5 failed the first two levels.

Why it matters

Standardized exams are not real jobs, but the speed of improvement is staggering. As models master technical finance knowledge, human value shifts toward judgment, trust, and client relationships.

The Deets

• 980 questions across all CFA tiers
• Constructed-response sections now scoring above 90%
• Performance jumped from 70s to near-perfect in ~24 months

Key takeaway

The floor for financial knowledge is rising fast.

🧩 Jargon Buster: Constructed Response - Exam questions requiring written answers instead of multiple choice.

Source: The Rundown AI

🛠️ Tools & Products

• Nemotron 3 - NVIDIA’s open-source agent model family designed for scalable multi-agent systems.
• Cursor Visual Design Editor - A new way to edit frontend designs directly inside Cursor without hopping to Figma.
• Quadratic - An AI-powered spreadsheet that turns raw data into insights and visuals.
• ManyPI - Converts any website into a clean, type-safe API for RAG and data extraction.

⚡ Quick Hits

• Chai Discovery raised $130M Series B at a $1.3B valuation to advance AI-driven drug design.
• Merriam-Webster named “slop” its 2025 Word of the Year, defining low-quality AI-generated content at scale.
• Amazon added “Ask This Book” to Kindle, answering spoiler-free questions with no author opt-out.
• British Airways warned airlines must optimize for AI agents or risk invisibility in travel search.

Today’s Sources: AI Secret, The Rundown AI, There’s An AI For That