By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Masters of Scale

Now is the time to go "all in" on AI (Microsoft AI's Mustafa Suleyman & Reid Hoffman) | Summit 2024

Artificial intelligence is a “new digital species,” says Mustafa Suleyman, Microsoft AI’s CEO. For this episode, Mustafa joined Reid Hoffman on stage at the October 2024 Masters of Scale Summit.

Reid HoffmanhostMustafa Suleymanguest

Nov 2, 202425mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

AI as a New Digital Species
Reid Hoffman asks whether the species metaphor is a useful lens for AI. Mustafa Suleyman explains that each new technology wave forces us to reach for metaphors, and that he proposes the digital species because it is the closest corollary and frames containment correctly.
- •Each new technology wave requires a metaphor, as with electricity or the transatlantic telephone
- •The species metaphor fits models that see, hear, and take actions on your behalf
- •The frame also clarifies what we don't want AI to become
- •Reid Hoffman points to Suleyman's book The Coming Wave
Risks to Watch and the Upside of Creative, Patient AI
Suleyman reframes hallucination as creativity and names two capabilities to be cautious about: recursive self-improvement and full autonomy. He then describes the upside, including AI that is non-judgmental and a study in which chatbot conversations reduced belief in conspiracy theories.
- •Hallucinations are the upside of models that give varied, malleable responses
- •Recursive self-improvement without human oversight may be tried in 2025
- •Autonomy in arbitrary digital environments raises risk
- •Well-designed AI need not be snarky or judgmental, though some designers will choose otherwise
- •A chatbot's patience reduced conspiracy beliefs in a paper he read
Why EQ Matters as Much as IQ
Hoffman asks about the founding principle at Inflection that emotional intelligence is as important as IQ. Suleyman says AI researchers tended to neglect the delivery vehicle, and that tone, style, and personality may matter more to consumers than raw facts.
- •IQ covers accuracy, speed, comprehensiveness, and relevancy
- •Researchers often neglect how information is delivered
- •Tone, style, and asking questions matter to most consumers
- •The agentic future is also about engineering personality
The Agentic Future: Seeing, Acting, and the Action Quotient
Hoffman asks how agents will play a role in the next two to five years. Suleyman says a companion must first see what you see, which enables ambiguous references and then actions, but reliability is still a few years away, as with the long path of voice recognition to 99.5% accuracy.
- •A companion that sees your screen enables references like 'that thing I saw'
- •Actions include browsing, using APIs, booking, buying, and planning
- •Current demos are flaky, comparable to LLMs before GPT-3
- •Getting from 50-60% to 99% accuracy is the challenge
- •Voice recognition took 15-20 years to cross the threshold
Voice Interfaces Unlock New Ways of Interacting
Discussing voice input and generative AI, Suleyman argues that interface shape governs what you can put in. Search boxes taught us to compress ideas into a few words, while voice lets us speak in full sentences and ask for things we never digitized, lowering the barrier to getting things done.
- •The search box taught people to speak in compressed queries
- •Voice allows full sentences, self-correction, and rambling
- •Models can respond in paragraphs, prompting new kinds of questions
- •Zero marginal cost and less friction mean people will ask AI for tasks they don't do today
Creativity, Lower Barriers, and Permanent Memory
Hoffman asks how agents will make us more creative. Suleyman says many fleeting thoughts never get voiced because typing is a high bar, and he predicts memory will be solved in 2025, making the AI a second brain that remembers prior conversations.
- •Sub-linguistic daily questions often go unexpressed because of effort
- •He searches only five to eight times a day because of the effort
- •He is confident permanent memory will be done in 2025
- •Memory extends the personal knowledge graph with documents, email, and calendar
- •The AI becomes an extension of your mind, which is why EQ matters
Model Scaling and Distillation
Hoffman turns to tactical questions for entrepreneurs. Suleyman says models are getting both bigger and smaller, describes distillation, and expects scale to keep delivering for two to three years while the focus moves to trajectories of actions across digital surfaces.
- •Large expensive models teach small models through distillation
- •Reinforcement learning from AI feedback appears to work well
- •He sees no slowdown in scale benefits for two to three years
- •New modalities include video and image
- •The hard part is action trajectories across browser, desktop, and phone
Prompts, Data, and Fine-Tuning Opportunities for Startups
Hoffman asks where new data will come from and how entrepreneurs should think about it. Suleyman distinguishes a question from a prompt, treats a prompt as high-quality data, and says fine-tuning on tens of thousands of good examples leaves room for startups in niche verticals.
- •A question differs from a prompt such as a multi-page style guide
- •Prompts act as high-quality instructions that steer a pre-trained model
- •Nuance and brand adherence need tens of thousands of good examples
- •Fine-tuning gives more stable adherence to a behavior policy
- •Niche verticals can access this data, creating room for startups
Small Models and the Ambient Sensing Future
Asked about small models, Suleyman says small is the future because large models activate many irrelevant pathways. He imagines knowledge compressed into tiny models on fridge magnets, earbuds, wearables, or plant pots, and says a two-person team could explore this now.
- •Frontier models light up billions of pathways irrelevant to a query
- •Tight use cases can run on much smaller, cheaper models
- •A fridge magnet model might need only a few tens of millions of parameters
- •The ambient sensing revolution could come alive through such devices
- •Few are pushing this yet, though it is feasible for small teams
Designing a More Human Future
Hoffman poses a closing question about what attendees should think about over the next two days and offers his own: how technologists design a more human future. He argues that we evolve our humanity through our technology, so design should look forward as well as back.
- •Being human is often framed by the past thousands of years
- •Technology, from mugs to podcasting equipment, changes who we are
- •The question is how emotions and compassion are expressed in our stance toward technology
Suleyman's Closing Challenge: Are You All In?
Suleyman asks the audience whether they are all in, calling this a transition moment comparable to past big technology shifts. He urges attention from founders, academics, and activists, says nothing is predetermined, and calls the moment both empowering and a great responsibility.
- •Past technology transitions reshaped the structure of things
- •It is a moment to found and scale companies or pivot careers
- •By 2050 the train will have left the station
- •Nothing is predetermined and we can shape AI for humanity's best
- •Being alive now is lucky, empowering, and a great responsibility

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.