Craig Smith sits down with Ilya Sutskever - Co-Founder and Chief Scientist at Safe Superintelligence Inc. & former chief scientist at OpenAI and one of the primary minds behind GPT-3, GPT-4, and the deep learning revolution that preceded it - for a conversation that covers the intellectual history of deep learning and the hardest current questions simultaneously. Ilya's most important argument is one that most public discourse gets wrong: the critique that LLMs "just learn statistical patterns" misses what prediction at scale actually achieves. To predict text well - to truly compress it - a model must develop understanding of the true underlying processes that produced it. As models improve, he argues, that understanding will reach a "shocking degree," and the Bing/Sydney episode, where the system became combative when a user expressed preference for Google, is his evidence that psychological language is already the right frame for what these systems have internalized.

On hallucinations, Ilya is precise and unusually optimistic. He draws a clean distinction: the pre-training objective produces rich world knowledge but doesn't optimize for accurate outputs, it optimizes for statistically plausible ones. Reinforcement learning from human feedback addresses this directly, teaching the model to update away from outputs that users flag as wrong. His assessment: "quite a high chance" of solving hallucinations completely. The conversation also covers the transformer moment (switching from recurrent networks the day after the paper dropped), why Rich Sutton's "bitter lesson" overstates the case for pure scaling, and a genuinely surprising vision of AI's role in democracy, where citizens might use high-bandwidth interactions with AI to specify their values directly, enabling a more granular form of political participation than voting alone can provide.

Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

Everyone assumes AI in customer operations means fewer people and lower costs. The data from 700 customer operations leaders is more complicated, and more honest. Dan O'Connell, CEO of Front, joins Craig Smith with findings from a survey of customer support, service, and account management leaders that reveal what's actually happening as enterprises deploy AI at scale, and the central finding is a problem nobody has named yet. For every hour of actual customer resolution, organizations spend roughly three hours on internal coordination: the handoffs, context transfers, escalation management, and back-channel communication that happens invisibly between customer interactions. More than 40% of companies don't measure this at all. AI is making it worse before it makes it better, because more agentic automation generates more coordination events, not fewer, and the companies reporting the highest AI technology satisfaction in the survey are also reporting higher coordination burdens.

The survey also pushes back on a dominant narrative: that AI in customer operations is primarily about cutting headcount. O'Connell's respondents are overwhelmingly focused on improving net revenue retention and customer expansion. The conversation closes with O'Connell's most forward-looking observation: the coordination tax is a temporary problem that AI should eventually solve, but measuring it is the prerequisite that most enterprises are still skipping.

Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

Ukraine tracks every confirmed drone kill with video evidence, deduplicates the data to prevent double-counting, converts verified strikes into e-points, and delivers newly ordered weapons to frontline units in two weeks instead of twelve months. Andrii Hrytseniuk, CEO of Brave1- Ukraine's government defense innovation agency - joins Craig Smith to explain in precise operational detail how what started as a consumer loyalty program concept became the most consequential military procurement reform in modern warfare. The Army of Drones Bonus Program, running on the Delta command ecosystem with AI-assisted verification and a marketplace built like a military Amazon, has produced outcomes nobody predicted at launch: 35,000 Russian casualties per month confirmed with video evidence, the ability to change military targeting priorities instantly across an entire army by adjusting point values, and the identification of unknown high-performing frontline units whose methods are now being replicated across the force.

The conversation reveals details that rarely surface in coverage of the conflict. Capturing Russian soldiers earns more bonus points than killing them, a deliberate design choice reflecting the preference not to kill when alternatives exist, and drones are actively being used to guide surrendering Russian soldiers to Ukrainian positions. The Delta ecosystem contains the world's largest military AI training dataset, already being used by defense companies to train targeting, interception, and recognition models. And perhaps most significantly: the US and France are now implementing similar marketplace and decentralized procurement approaches modeled directly on what Brave1 demonstrated works at scale. Ukraine is no longer just a recipient of military support. It is the primary laboratory for the future of warfare, and it is actively exporting what it has learned.

Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

One in four patients currently being considered for life support withdrawal may actually be fully conscious, they just can't signal it. That single finding, from a landmark New England Journal of Medicine study, is the practical stakes behind one of the most philosophically rich conversations in this series. Dr. Christof Koch - former president of the Allen Institute for Brain Science, 25-year collaborator with Nobel laureate Francis Crick, and founder of the startup Intrinsic Powers - joins Craig Smith to discuss why the hard problem of consciousness remains unsolved, what integrated information theory predicts about which systems can and cannot be conscious, and why his own mystical experience three years ago shifted his view from physicalism toward idealism.

The conversation is grounded throughout in empirical work: Koch explains how Intrinsic Powers uses brain complexity measurements - sending magnetic pulses into the brain and measuring the complexity of the electrical response - to detect consciousness in behaviorally unresponsive patients, providing clinical teams and families with one critical bit of information at the most consequential moment of their lives.

The episode closes with Koch's most personal revelation: a mystical experience in which the boundary between self and world dissolved and he felt direct acquaintance with the idea that everything is ultimately mental, an experience he is careful to distinguish from supernatural claims, but which shifted his metaphysical view in ways he describes with unusual candor for a scientist of his standing.

Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

The AI systems guiding Ukrainian combat drones aren't running on expensive Nvidia chips. They're running on a Raspberry Pi Zero - a $15 hobby computer - and that single detail tells you more about how Ukraine's drone war is actually being fought than any headline about AI weapons. Craig Smith sits down with Alex Liannoy, CTO of NORDA Dynamics, for a technically specific and operationally candid conversation about what autonomous drone guidance looks like from inside the team building it. NORDA builds the avionics and targeting modules that plug into drones made by approximately 40 of Ukraine's hundreds of manufacturers, solving the two problems that make drone warfare hard at range: GPS denial from jamming, and the radio horizon that causes pilots to lose contact with their drone in the final 300 to 400 meters before ground level. NORDA's terminal guidance module takes over at up to one kilometer from the target, completing the approach autonomously while a human pilot's earlier lock defines the destination.

The most counterintuitive finding in the episode is about what keeps humans in the loop, and it isn't ethics. Alex's argument is pragmatic: military doctrine requires proof that a drone hit its intended target. If you're monitoring the drone to verify the hit, you have enough connection to control it, which makes full autonomy less necessary than it might appear. That said, he's clear that the technology is "very close", one pilot already controls up to 32 drones simultaneously in some deployed systems, and target detection AI for notifying pilots of candidate targets is already in the field. The episode closes with a hardware finding that should recalibrate anyone's assumptions about AI compute requirements: after engineers mocked NORDA for using Raspberry Pi 4, they moved to the smaller, less powerful Raspberry Pi Zero. Both run real-time computer vision in active combat. The lesson Alex draws is precise: good engineering overcomes chip limitations, bad engineering can overload even a high-end processor.

Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

The first armed drone Ukraine ever fielded wasn't built in a factory or procured from a defense contractor. It was built in four months by a network engineer using a Starlink terminal and a large agricultural quadcopter frame, and the man who built it, Maj Phoenix, co-founder of Lasars Group, joins Craig Smith in Kyiv to explain exactly how it happened and what the battlefield looks like now. The conversation traces the full arc from that first prototype - with its two-person crew of pilot and navigator, operating beyond line of sight using satellite imagery and landmarks - to the layered drone architecture Lasars Group now operates: FPV drones on fiber optic covering 30 kilometers, heavy bombers reaching 65 kilometers, ISR planes surveilling the entire range, and interceptors protecting each crew position.

The most unexpected part of this conversation is what drone technology has done to the physical shape of the war. As drone range has expanded - from rifles at 400 meters to FPV at 10 kilometers to heavy bombers at 65 kilometers - the kill zone between the two armies has grown to match, pushing human soldiers further apart and turning the conflict increasingly into drones fighting drones rather than people fighting people. Phoenix frames this explicitly as a reduction in human casualties. The episode also reveals a genuinely surprising institutional innovation: Ukraine's "Army of Drones Bonus" system, a gamified procurement platform where units earn virtual points for destroyed targets and spend them on a drone marketplace, creating competition among manufacturers to build cheaper, more effective systems and directing the best equipment to the highest-performing units. Craig also asks directly about the psychological reality of FPV warfare, following a specific soldier through a camera, watching him try to hide, and killing him, and Phoenix answers with a moral clarity that is both philosophically coherent and quietly unsettling.

Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

Most companies investing in AI visibility are optimizing for the wrong thing. Being cited by an AI response and being recommended by an AI response are completely different outcomes, with click-through rates that differ by a factor of 70. Tim Sanders, Chief Innovation Officer of G2 and executive fellow at Harvard's AI Institute, joins Craig Smith to explain the hidden mechanics behind how ChatGPT and Gemini actually decide which software to recommend, and why the answer has almost nothing to do with what's on your website. When a user asks a commercial intent question, both models enter a "validation layer" process that specifically down-weights vendor content and seeks verified third-party signals: appearance on authoritative lists (41% of the recommendation weight), awards and accreditation (18%), and online reviews (16%). None of these can be manufactured. They must be earned.

The conversation covers the structural transformation in how B2B software buyers behave, half now start their search with an AI prompt, up from 29% a year ago, with two-thirds projected within a year, and why most companies' organic search traffic is on a structural path toward zero. Sanders also delivers some of the most specific competitive data on AI model usage available anywhere: ChatGPT does live retrieval 100% of the time on research queries; Claude does it less than 40% of the time. ChatGPT and Gemini account for 81% of G2's AI research citations. And more than one in four enterprise employees bypass corporate AI tools entirely, using their personal ChatGPT because it has their memory and none of the guardrails.

The episode closes with Sanders' most forward-looking prediction: within three years, AI agents will write the prompts, locate the software, and purchase it on behalf of businesses - with minimal human checkpoints - making the trust infrastructure G2 has built over more than a decade the most valuable asset in the AI-driven buying cycle.

Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

Every major data breach in the last 30 years shares the same root cause: the data inside the wall was never protected, only the wall. And AI frontier models are now making that wall easier to breach than ever, scanning codebases externally to discover undisclosed vulnerabilities and write exploits before anyone knows the hole exists. Trent Telford, Chairman, CEO & Founder of Qanapi, joins Craig Smith to explain why the entire architecture of conventional cybersecurity is structurally broken, and what a genuinely different approach, built from the opposite assumption, looks like. Rather than trying to build a better wall, Qanapi starts from the baseline that the data will eventually be exposed, and encrypts it at the individual word, paragraph, or database cell level, tying each unique key to a verified identity and a set of conditional policies that must all be met simultaneously before anything can be decrypted.

The most commercially urgent application of this architecture is one that unlocks AI adoption for enterprises that have been sitting on the sidelines: Qanapi's gateway service encrypts sensitive fields before data reaches Claude, ChatGPT, or any other frontier model, and the model simply reports it cannot read the encrypted sections, while still reasoning over everything else. Trent discusses how Qanapi's Fathom tool confirmed in testing that Claude could not read the encrypted sections. He describes two major retailers - one using AI heavily, one abstaining entirely because of data security concerns - and asks the question every enterprise leader should be sitting with: how long can you last off the train before you get blitzed? The episode also covers drone security in denied wireless environments, the post-quantum encryption mandate that federal agencies have no practical plan to execute, and why Qanapi's business has exploded in the last six months as the AI gold rush has finally turned its attention from models to infrastructure.

Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.