AI Leaders Reveal the Next Wave of AI Breakthroughs (At FII Miami 2025) | EP #150
A live panel at the Future Investment Initiative (FII) Priority Summit in Miami, moderated by Peter Diamandis, brings together CEOs from across the AI stack: Prem Akkaraju (Stability AI) on generative media reaching photorealism, Ramin Hasani (Liquid AI) on ultra-efficient liquid neural networks running privately on-device, Jack Hidary (SandboxAQ) on quantitative AI for drug discovery, materials, and GPS-denial defense, Jim Keller (Tenstorrent) on cheaper, open, tensor-native AI hardware, and Alexander Sukharevsky (McKinsey QuantumBlack) on why most enterprise AI pilots never reach production. Recurring threads include the gap between AI hype and deployed value (11% and 7% production success rates), the case for small, empowered teams over massive top-down programs, and the shift from language-only models toward quantitative and physically-grounded AI. The panel closes on a shared call to lean into AI adoption rather than resist it, framing the next five years as decisive for companies, industries, and nations.
Stability AI's step-by-step professional content pipelineAI▶ 0:53
Prem Akkaraju explains Stability AI (creator of Stable Diffusion) is fine-tuning dozens of 'ultra narrow' AI models for individual stages of professional film/TV production (rig removal, roto, camera match, compositing) rather than relying on a single end-to-end text-to-video model.
Timeline to photorealistic, on-demand video generationAI▶ 2:57
Diamandis asks how close AI is to generating indistinguishable-from-real video on demand; Akkaraju says some workflows are already there and full on-the-fly photorealistic generation is likely within 6-12 months.
Liquid AI's liquid neural networks and private, on-device AIAI▶ 3:56
Ramin Hasani describes Liquid AI's non-Transformer architecture (invented at MIT/Daniela Rus's lab), which needs minimal compute, enabling ChatGPT-like experiences to run locally on phones, laptops, and other devices for near-zero cost and full data privacy.
Liquid neural networks flying a US Air Force fighter jetAI▶ 5:34
Hasani notes Liquid AI's architecture was trusted by the US Air Force as the first neural network to autonomously navigate a fighter jet, illustrating private/on-device AI extending to cars, satellites, and jets.
SandboxAQ's large quantitative models (LQMs) for drug and materials discoveryBiotech▶ 8:01
Jack Hidary contrasts commodity large language models with SandboxAQ's 'large quantitative models' trained on molecules and atoms (not text), which convert quantum physics equations into GPU-friendly matrix algebra to design new drugs and materials.
Quantum computing roadmap and GPU-QPU hybrid futureCompute▶ 10:59
Hidary discusses using classical GPUs today to approximate quantum equations, with real quantum computers expected to mature over roughly 5-7 years and eventually merge into a hybrid GPU/QPU cloud mesh.
Aramco partnership: converting hydrocarbons into advanced materialsEnergy▶ 11:29
Hidary cites Aramco as SandboxAQ's newest customer, using quantitative AI to convert raw hydrocarbons into higher-order carbon-hydrogen composites for lighter cars, rockets, and aircraft, rather than low-grade fuels.
Tenstorrent's tensor-native, open hardware and software stackCompute▶ 12:44
Jim Keller explains Tenstorrent builds native tensor processors (simpler to program than GPUs, natively networked to each other) and has open-sourced its software stack, aiming to make AI hardware dramatically cheaper and less proprietary.
Following the open-source DeepSeek R1 release, Keller argues the open-source AI landscape is a 'mixed bag' (research and some weights open, infrastructure mostly not) and commits to open-sourcing Tenstorrent's full software and hardware stack to avoid AI control concentrating among a few large players.
The AI production gap: pilots vs. real deploymentEconomy▶ 17:24
Alexander Sukharevsky of McKinsey's QuantumBlack presents data showing only about 11% of enterprise AI use cases over the past 5 years reached production, dropping to roughly 7% for generative AI specifically, framing QuantumBlack's mission as closing that gap.
From digital transformation failure to an 'age of creativity'Economy▶ 19:04
Sukharevsky argues most AI/digital transformations fail because companies force new technology onto old, broken processes instead of reinventing them, and frames the moment as the end of an 'age of mediocrity' and the start of an 'age of creativity' as commodity work is automated.
Small teams and 'moonshot' units solving hard problemsAI▶ 24:53
Jack Hidary advocates for small, empowered teams armed with LLMs/LQMs, citing an 11-person SandboxAQ team that solved GPS jamming/spoofing for the US Air Force, and urges leaders to carve out small mission-driven teams inside larger organizations.
GPS jamming and spoofing as a live geopolitical/aviation problemGeopolitics▶ 25:19
Hidary describes GPS being actively jammed or spoofed over parts of Europe, the Gulf region, and the Indo-Pacific (attributed to PRC actions), and how SandboxAQ's quantum sensor technology now flies operationally with the US Air Force as a fix.
Predictions made
openPrem Akkaraju: Fully photorealistic, on-demand video generation (indistinguishable from reality) will be achievable.
EP #? · · due: Within 6-12 months of Feb 2025 (roughly by early-to-mid 2026) · ▶ watch
“I would say we're probably... that's going to happen this year, I think within six, 12 months.”
Your call:
openPrem Akkaraju: Stability AI's tools will be used in the upcoming Avatar sequels (Avatar 3, 4, and 5).
EP #? · · due: Avatar 3 targeted for December (2025) · ▶ watch
“We're going to see it, we're going to see it in Avatar 3, 4 and 5... hopefully it'll come out in December.”
Your call:
openJack Hidary: Quantum computing hardware will mature to the point of forming a hybrid GPU/QPU (Quantum Processing Unit) cloud mesh.
EP #? · · due: About 5 to 7 years from Feb 2025 (roughly 2030-2032) · ▶ watch
“These announcements you're going to see come in a great cadence, culminating in a crescendo... in about 5 to seven years of having great quantum computers. We'll add those to the arsenal, we'll have GPU, QPU... in one mesh, cloud hybrid.”
Your call:
openPeter Diamandis: By the end of this decade there will be two kinds of companies: those fully utilizing AI, and those that are out of business.
“By the end of this decade there['re] going to be two kinds of companies: those that are fully utilizing AI, and those that are out of business.”
Your call:
Numbers that matter
270 million downloadsCumulative downloads of Stable Diffusion to date, per Prem Akkaraju.
9 million downloadsDownloads of the next most popular image AI model, cited for comparison to Stable Diffusion's 270 million.
$2 billion valuation in ~2 yearsLiquid AI's valuation growth since founding, per Peter Diamandis's introduction of Ramin Hasani.
quarter of a billion dollar funding roundLiquid AI's recent raise led by g42.
$850 millionSandboxAQ's funding raised, described by Jack Hidary as their seed round.
5 to 7 yearsJack Hidary's estimated timeline for mature quantum computers to join GPUs in a hybrid compute mesh.
$700 million Series DTenstorrent's recent funding round, noted by Peter Diamandis.
millions to trillions of trillions of instructions per secondJim Keller on the scale-up of computing power over his 40-year career.
600 lines of codeSize of Tenstorrent's top-level software stack for running large models, per Jim Keller.
5,000 people across 50 countriesSize and global reach of McKinsey's QuantumBlack AI team, per Alexander Sukharevsky.
11%Share of enterprise AI use cases over the last 5 years that reached production, per Alexander Sukharevsky.
7%Success rate specifically for generative AI use cases reaching production, per Alexander Sukharevsky.
5 R&D centers, 43 deployed productsQuantumBlack's global research footprint and number of products deployed, per Alexander Sukharevsky.
98% of films made on digital by 2017Prem Akkaraju's analogy for technology adoption curves, citing the film industry's shift from film to digital starting in 2000.
11-person teamSize of the SandboxAQ team that solved GPS jamming/spoofing now operational with the US Air Force, per Jack Hidary.
Worth digging into
🕳️ Liquid AI's liquid neural networks autonomously navigating a US Air Force fighter jet
A striking, specific military-AI claim (first neural net trusted to fly a fighter jet) with no named program or verification given on stage.
🕳️ SandboxAQ's 11-person team fixing GPS jamming/spoofing now flying with the US Air Force
A concrete, checkable operational deployment (quantum sensors + AI) tied to a live geopolitical problem (GPS denial in Europe, the Gulf, and the Indo-Pacific).
🕳️ McKinsey QuantumBlack's 11% / 7% AI production success-rate statistics
These numbers are cited as headline evidence of an AI hype-vs-deployment gap but no source study or methodology is given.
🕳️ SandboxAQ's Aramco partnership converting hydrocarbons into carbon composites
A concrete industrial application of 'large quantitative models' with potential relevance to materials science and clean(er) heavy industry.
🕳️ Tenstorrent's fully open-sourced hardware and software stack
Jim Keller frames this as a democratization strategy and also as their best hiring channel, which is an interesting recruiting/open-source case study.