reSee.it - Related Video Feed

"It's actually the biggest misconception." "We're not designing them." "First fifty years of AI research, we did design them." "Somebody actually explicitly programmed this decision, previous expert system." "Today, we create a model for self learning." "We give it all the data, as much compute as we can buy, and we see what happens." "We kinda grow this alien plant and see what fruit it bears." "We study it later for months and see, oh, it can do this." "It has this capability." "We miss some." "We still discover new capabilities and old models." "Or if I prompt it this way, if I give it a tip and threaten it, it does much better." "But, there is very little design."

We will become a hybrid species, still human but enhanced by AI, no longer limited by our biology, and free to live life without limits. We're going to find solutions to diseases and aging. Having worked in AI for sixty-one years, longer than anyone else alive, and being named one of Time's 100 most influential people in AI, I predicted computers would reach human-level intelligence by 2029, and some say it will happen even sooner.

"It's really weird to, like, live through watching the world speed up so much." "A kid born today will never be smarter than AI ever." "A kid born today, by the time that kid, like, kinda understands the way the world works, will just always be used to an incredibly fast rate of things improving and discovering new science." "They'll just they will never know any other world." "It will seem totally natural." "It will seem unthinkable and stone age like that we used to use computers or phones or any kind of technology that was not way smarter than we were." "You know we will think like how bad those people of the 2020s had it."

Amy and her colleague discuss integrating AI-native innovation with a human-centered design approach, focusing on how technology can be made accessible through natural interaction with AI and through rapid, user-friendly development flows. They begin by positioning AI as the new user interface. The other speaker notes that AI’s ease and approachability come from the ability to use human language, enabling conversations that let people interact with technology in a fundamentally new way. This language-based interaction is highlighted as a core shift in how users engage with digital tools and services. Beyond language, the conversation expands to include other modalities that users can employ to communicate with AI. The speakers identify text, images, and audio as essential inputs. The concept of multimodality is introduced to describe the ability to input using whatever format feels most natural to the user. Examples given include dropping in a screenshot, using voice to talk to the AI, or providing a video or a document. The emphasis is on a flexible, conversational experience that can accept diverse media and still deliver the necessary answers and help. The speakers then pivot to the question of how to create applications quickly and easily. They express enthusiastic interest in a partnership with Figma, a design platform. The collaboration is described as enabling designers who create an application design in Figma to hand off that design to a build agent, which can translate the design into an enterprise-grade application. This suggests a streamlined pipeline from design to production, leveraging AI to automate aspects of the development process and accelerate delivery while maintaining enterprise quality. Throughout, the emphasis remains on combining AI-driven capabilities with human-centered design principles to simplify interactions and speed up application development. The dialogue underscores the idea that users can engage with AI through natural language and multiple input formats, and that design-to-deployment workflows can be accelerated through integrated tools and partnerships. To learn more about AI experience, the conversation points listeners to a link in the comments, inviting further exploration of the described capabilities and partnerships.

"We are at the point where we can create very believable, realistic virtual environments." "We're also getting close to creating intelligent agents." "If you just take those two technologies and you project it forward and you think they will be affordable one day, a normal person like me or you can run thousands, billions of simulations." "Then those intelligent agents, possibly conscious ones, will most likely be in one of those virtual worlds, not in the real world." "In fact, I can, again, retro causally place you in one." "I can commit right now to run billion simulations of this exact interview." "Mhmm. So the chances are you're probably in one of those." "One, we don't know what resources are outside of the simulation. This could be like a cell phone level of compute."

Everybody's an author now. Everybody's a programmer now. That is all true. And so we know that AI is a great equalizer. We also know that, it's not likely that although everybody's job will be different as a result of AI, everybody's jobs will be different. Some jobs will be obsolete, but many jobs will be created. The one thing that we know for certain is that if you're not using AI, you're going to lose your job to somebody who uses AI. That I think we know for certain. There's not

Former Tesla AI director Andre Karpathy discusses software in the era of AI, emphasizing how software is changing at a fundamental level and what this means for students entering the industry. Key framework: three generations of software - Software 1.0: the code that programs computers. - Software 2.0: neural networks, where you tune data sets and run optimizers to create model parameters; the weights program the neural nets rather than hand-written code. - Software 3.0: prompts as programs that program large language models (LLMs); prompts are written in English, effectively a new programming language. - He notes that a growing amount of GitHub-like activity in software 2.0 blends English with code, and that the ecosystem around LLMs resembles a newer GitHub-like space (e.g., Hugging Face, Model Atlas). An example: tuning a LoRa on Flux’s image generator creates a “git commit” in this space. Evolving software stacks in practice - At Tesla Autopilot, the stack evolved from heavy C++ (software 1.0) to neural nets handling image processing and sensor fusion, with many 1.0 components being migrated to 2.0. The neural network grew in capability and size, and the 1.0 code was deleted as functionality migrated to 2.0. - We now have three distinct programming paradigms: 1.0 coding, 2.0 weights, and 3.0 prompts. Fluent capability in all three is valuable because tasks may be best solved with code, trained networks, or prompts. LLMs as a new computer and ecosystem view - Andrew Ng’s “AI is the new electricity” is cited to frame LLMs as utility-like (CapEx for training, OpEx for API serving, metered usage, low latency, high uptime) and also as fabs-like (large CapEx, rapid tech-tree growth), though software nature means more malleability. - LLMs are compared to operating systems: CPU-like core, memory in context windows, and orchestration of compute/memory for problem solving. App downloads can be run across various LLM platforms similarly to cross-OS apps. - The diffusion pattern of LLMs is inverted compared to many technologies: governments and corporations often lag behind consumer adoption, with AI topics sometimes used for everyday tasks like “boiling an egg” rather than high-level strategic aims. Practical implications for developers and students - Build fluently across paradigms: code in 1.0, tune 2.0 models, and design 3.0 prompts; decide when to code, train, or prompt depending on task. - Partially autonomous apps: exemplified by Cursor and Perplexity. - Cursor: traditional interface plus LLM integration, with under-the-hood embeddings, diffs, and multi-LLM orchestration; GUI support for auditing changes; autonomy slider lets users control how much the AI acts vs. what humans verify. - Perplexity: similar features, with sources cited and ability to scale autonomy from quick search to deep research. - Autonomy slider concept: users can limit or increase AI autonomy depending on task complexity; the AI handles context management and multi-call orchestration, while humans verify for correctness and security. - Education and “keeping AI on the leash”: emphasize concrete prompts, better verification, and development of structured education pipelines with auditable AI-generated content. Opportunities and caveats in AI-assisted workflows - Education and governance: separate roles for AI-generated courses and AI-assisted delivery to students, ensuring syllabus adherence and auditability. - Documentation and access for LLMs: docs should be machine-readable (e.g., markdown), and wording should be actionable (avoid “click” commands; provide equivalent API calls like curl) to facilitate LLM interactions. - Tools to ingest data for LLMs: services that convert GitHub repos into ingestible formats (e.g., git ingest, DeepWiki) to create ready-to-query knowledge bases. - Agents vs. augmentation: early emphasis on augmentation (Iron Man-like suits) rather than fully autonomous systems; the autonomy slider enables gradual handover from human supervision to more autonomous tasks while maintaining safety and auditability. - The future of “native” programming: vibe coding and byte coding illustrate how language-based programming lowers barriers, enabling broad participation in software creation; the takeaway is that natural-language interfaces can act as a gateway to software development, even for non-experts. Closing synthesis - We’re at an era where enormous code rewriting is needed, and LLMs function as utilities, fabs, and operating systems, though still early—like the 1960s of OS development. - The next decade will likely feature a spectrum of partially autonomous products with specialized GUIs and rapid verification loops, guided by an autonomy slider and careful human oversight. - Karpathy envisions an ongoing collaboration with AI: building partial autonomy products, evolving tooling, and experimenting with how the industry and education adapt to this new programming reality. He invites readers to participate in shaping this future.

"You know, in the near future, we're all going to be working around with AI assistance, helping us in our daily lives that we're going to be able to interact with through various smart devices including smart glasses and things like that, through voice and through various other ways of interacting with them." "So, I have smart glasses with cameras and displays in them, etcetera." "Currently, you can have smart glasses without displays, but soon the displays will exist." "Right now they exist." "They're just too expensive to be commercialized." "This is the Orion demonstration built by our colleagues at Meta." "So, future is coming and the vision is that all of us will be basically working around with AI assistants all our lives." "It's like all of us will be kind of like a high level CEO or politician or something, running around with a staff of smart virtual people working for us." "That's kind of the possible picture."

- The conversation centers on how AI progress has evolved over the last few years, what is surprising, and what the near future might look like in terms of capabilities, diffusion, and economic impact. - Big picture of progress - Speaker 1 argues that the underlying exponential progression of AI tech has followed expectations, with models advancing from “smart high school student” to “smart college student” to capabilities approaching PhD/professional levels, and code-related tasks extending beyond that frontier. The pace is roughly as anticipated, with some variance in direction for specific tasks. - The most surprising aspect, per Speaker 1, is the lack of public recognition of how close we are to the end of the exponential growth curve. He notes that public discourse remains focused on political controversies while the technology is approaching a phase where the exponential growth tapers or ends. - What “the exponential” looks like now - There is a shared hypothesis dating back to 2017 (the big blob of compute hypothesis) that what matters most for progress are a small handful of factors: compute, data quantity, data quality/distribution, training duration, scalable objective functions, and normalization/conditioning for stability. - Pretraining scaling has continued to yield gains, and now RL shows a similar pattern: pretraining followed by RL phases can scale with long-term training data and objectives. Tasks like math contests have shown log-linear improvements with training time in RL, and this pattern mirrors pretraining. - The discussion emphasizes that RL and pretraining are not fundamentally different in their relation to scaling; RL is seen as an RL-like extension atop the same scaling principles already observed in pretraining. - On the nature of learning and generalization - There is debate about whether the best path to generalization is “human-like” learning (continual on-the-job learning) or large-scale pretraining plus RL. Speaker 1 argues the generalization observed in pretraining on massive, diverse data (e.g., Common Crawl) is what enables the broad capabilities, and RL similarly benefits from broad, varied data and tasks. - The in-context learning capacity is described as a form of short- to mid-term learning that sits between long-term human learning and evolution, suggesting a spectrum rather than a binary gap between AI learning and human learning. - On the end state and timeline to AGI-like capabilities - Speaker 1 expresses high confidence (~90% or higher) that within ten years we will reach capabilities where a country-of-geniuses-level model in a data center could handle end-to-end tasks (including coding) and generalize across many domains. He places a strong emphasis on timing: “one to three years” for on-the-job, end-to-end coding and related tasks; “three to five” or “five to ten” years for broader, high-ability AI integration into real work. - A central caution is the diffusion problem: even if the technology is advancing rapidly, the economic uptake and deployment into real-world tasks take time due to organizational, regulatory, and operational frictions. He envisions two overlapping fast exponential curves: one for model capability and one for diffusion into the economy, with the latter slower but still rapid compared with historical tech diffusion. - On coding and software engineering - The conversation explores whether the near-term future could see 90% or even 100% of coding tasks done by AI. Speaker 1 clarifies his forecast as a spectrum: - 90% of code written by models is already seen in some places. - 90% of end-to-end SWE tasks (including environment setup, testing, deployment, and even writing memos) might be handled by models; 100% is still a broader claim. - The distinction is between what can be automated now and the broader productivity impact across teams. Even with high automation, human roles in software design and project management may shift rather than disappear. - The value of coding-specific products like Claude Code is discussed as a result of internal experimentation becoming externally marketable; adoption is rapid in the coding domain, both internally and externally. - On product strategy and economics - The economics of frontier AI are discussed in depth. The industry is characterized as a few large players with steep compute needs and a dynamic where training costs grow rapidly while inference margins are substantial. This creates a cycle: training costs are enormous, but inference revenue plus margins can be significant; the industry’s profitability depends on accurately forecasting future demand for compute and managing investment in training versus inference. - The concept of a “country of geniuses in a data center” is used to describe the point at which frontier AI capabilities become so powerful that they unlock large-scale economic value. The timing is uncertain and depends on both technical progress and the diffusion of benefits through the economy. - There is a nuanced view on profitability: in a multi-firm equilibrium, each model may be profitable on its own, but the cost of training new models can outpace current profits if demand does not grow as fast as the compute investments. The balance is described in terms of a distribution where roughly half of compute is used for training and half for inference, with margins on inference driving profitability while training remains a cost center. - On governance, safety, and society - The conversation ventures into governance and international dynamics. The world may evolve toward an “AI governance architecture” with preemption or standard-setting at the federal level, to avoid an unhelpful patchwork of state laws. The idea is to establish standards for transparency, safety, and alignment while balancing innovation. - There is concern about autocracies and the potential for AI to exacerbate geopolitical tensions. The idea is that the post-AGI world may require new governance structures that preserve human freedoms, while enabling competitive but safe AI development. Speaker 1 contemplates scenarios in which authoritarian regimes could become destabilized by powerful AI-enabled information and privacy tools, though cautions that practical governance approaches would be required. - The role of philanthropy is acknowledged, but there is emphasis on endogenous growth and the dissemination of benefits globally. Building AI-enabled health, drug discovery, and other critical sectors in the developing world is seen as essential for broad distribution of AI benefits. - The role of safety tools and alignments - Anthropic’s approach to model governance includes a constitution-like framework for AI behavior, focusing on principles rather than just prohibitions. The idea is to train models to act according to high-level principles with guardrails, enabling better handling of edge cases and greater alignment with human values. - The constitution is viewed as an evolving set of guidelines that can be iterated within the company, compared across different organizations, and subject to broader societal input. This iterative approach is intended to improve alignment while preserving safety and corrigibility. - Specific topics and examples - Video editing and content workflows illustrate how an AI with long-context capabilities and computer-use ability could perform complex tasks, such as reviewing interviews, identifying where to edit, and generating a final cut with context-aware decisions. - There is a discussion of long-context capacity (from thousands of tokens to potentially millions) and the engineering challenges of serving such long contexts, including memory management and inference efficiency. The conversation stresses that these are engineering problems tied to system design rather than fundamental limits of the model’s capabilities. - Final outlook and strategy - The timeline for a country-of-geniuses in a data center is framed as potentially within one to three years for end-to-end on-the-job capabilities, and by 2028-2030 for broader societal diffusion and economic impact. The probability of reaching fundamental capabilities that enable trillions of dollars in revenue is asserted as high within the next decade, with 2030 as a plausible horizon. - There is ongoing emphasis on responsible scaling: the pace of compute expansion must be balanced with thoughtful investment and risk management to ensure long-term stability and safety. The broader vision includes global distribution of benefits, governance mechanisms that preserve civil liberties, and a cautious but optimistic expectation that AI progress will transform many sectors while requiring careful policy and institutional responses. - Mentions of concrete topics - Claude Code as a notable Anthropic product rising from internal use to external adoption. - The idea of a “collective intelligence” approach to shaping AI constitutions with input from multiple stakeholders, including potential future government-level processes. - The role of continual learning, model governance, and the interplay between technology progression and regulatory development. - The broader existential and geopolitical questions—how the world navigates diffusion, governance, and potential misalignment—are acknowledged as central to both policy and industry strategy. - In sum, the dialogue canvasses (a) the expected trajectory of AI progress and the surprising proximity to exponential endpoints, (b) how scaling, pretraining, and RL interact to yield generalization, (c) the practical timelines for on-the-job competencies and automation of complex professional tasks, (d) the economics of compute and the diffusion of frontier AI across the economy, (e) governance, safety, and the potential for a governance architecture (constitutions, preemption, and multi-stakeholder input), and (f) the strategic moves of Anthropic (including Claude Code) within this evolving landscape.

Andrej (Andre) began by noting his notable roles in building and explaining modern AI, including co-founding OpenAI, getting autopilot working at Tesla, and popularizing “vibe coding.” He described a surprising shift from December onward: he felt more behind as a programmer, then observed that the latest models began producing correct code chunks without edits, enabling a more coherent, end-to-end workflow and fueling countless side projects. On the idea that LLMs are a new computing paradigm, Andre explained software three point zero as programming via prompting, where the context window and the LLM act as an interpreter performing computation. He contrasted software one point zero (writing code) and software two point zero (curating data sets and training networks). He gave concrete illustrations: - OpenClaw: installation becomes a copy-paste task for an agent that configures and runs across environments, rather than writing a complex shell script. - MenuGen: photographing a restaurant menu to generate item images and descriptions; with Gemini, the agent could overlay items directly onto the menu image, rendering the final result without the traditional app scaffolding. This demonstrated how software three point zero can perform tasks that previous paradigms required separate apps for. He emphasized that this paradigm shift isn’t just about faster coding; it enables new capabilities in unstructured information processing, such as building a knowledge wiki from documents, where data can be reordered and reframed in novel ways. Looking ahead, Andre speculated about 2026-era web development, where neural computers could render UIs from raw inputs, and where neural networks could become the host process with CPUs acting as co-processors. He suggested that intelligence compute will dominate FLOPs, and that the progression could unfold in unexpected ways, with neural nets taking on primary workloads over time. On verifiability, Andre described how LLMs trained with reinforcement learning in verification-rich environments excel in verifiable domains like math and code, while remaining jagged in other areas. He cited examples like a car-wash decision (walk vs. drive) and a strawberry-letter-counting discrepancy, noting that model behavior can be robust in some tasks but flawed in others due to training data and RL focus. He warned that the labs’ data distributions shape capabilities, and emphasized that users must stay in the loop, fine-tune, or adapt when models operate outside familiar circuits. For founders, he argued verifiability remains a viable path: if a domain is verifiable and RL environments can be created, founders can still fine-tune and deploy effectively. He asserted that “everything is automatable” to some extent, though some tasks are easier than others. When discussing vibe coding vs. agentic engineering, Andre distinguished vibe coding as raising the floor for all software users, versus agentic engineering, which preserves quality standards and professional software discipline while enabling faster delivery through autonomous agents. He stressed a high ceiling for agentic engineers, noting that top performers can far exceed prior “10x” expectations. In terms of human skills, as agents assume more tasks, judgment, taste, oversight, and architectural design remain crucial. He illustrated this with MenuGen’s cross-domain edge cases (e.g., matching Stripe and Google email identities) where agents can falter, underscoring the need for human specification of plans and top-level coherence. On education, Andre closed with the idea that “you can outsource your thinking, but you can’t outsource your understanding.” He emphasized the enduring value of understanding and human direction, even as tools proliferate for synthetic data generation and knowledge bases. Overall, the conversation framed a shift to agent-native workflows, the critical role of verifiability and human oversight, and a vision of a future where neural networks largely drive perception, decision, and action, with humans guiding strategy and design.

And I I think that that AI, in my case, is creating jobs. It causes us to be able to create things that other people would customers would like to buy. It drives more growth. It drives more jobs. The other thing that that to remember is that AI is the greatest technology equalizer of all time.

Being surrounded by "superhuman" experts doesn't make one feel unnecessary; instead, it empowers confidence to tackle ambitious goals. Similarly, super AIs will empower people, making them feel confident. Using tools like Chat GPT increases feelings of empowerment and the ability to learn. AI reduces barriers to understanding almost any field, acting as a personal tutor available at all times. Everyone should acquire an AI tutor to teach them anything, including programming, writing, analysis, thinking, and reasoning, to feel more empowered.