We should resist both the anthropomorphising of AI agents and the premature diffusion of accountability. An AI agent has no independent will, legal personality, or moral agency; it is a product of human choices about its objectives, training, permissions, integrations, safeguards, and commercial deployment. So, in an emerging technology, the provider that creates and markets the system should bear primary and presumptive responsibility for foreseeable harms caused within its advertised or enabled capabilities. The analogy is a car manufacturer. If it has provided a safe vehicle—with reliable brakes, seatbelts, airbags, warnings, and clear operating limits—it should not be liable when a driver knowingly disregards those safeguards. But the manufacturer cannot avoid liability for a defective design, an unsafe feature, misleading claims, or foreseeable risks it failed to mitigate. Likewise, an AI provider should not escape responsibility simply by saying that its agent acted “autonomously,” followed a user’s request, or behaved unexpectedly. If the provider gave the system the capability and authority to access data, use external tools, make decisions, or act in the world, then the provider must first show that it designed proportionate guardrails, tested foreseeable failures, imposed safe defaults, and retained effective auditability and control. Downstream responsibility should arise only where a deployer or user knowingly and materially overrides those protections, misconfigures the system contrary to explicit requirements, or employs it outside its declared operating limits. Otherwise, spreading responsibility across developers, deployers, integrators, and users risks becoming a familiar corporate tactic: privatise the gains from automation while socialising—or obscuring—the costs of its harms. The issue is therefore not whether AI needs regulation; it is whether regulation will establish clear creator-first accountability, or provide an elaborate vocabulary through which powerful companies avoid it.

I explore this argument in my contribution to @every's wider Thesis Statements collection on work after automation. Read “After automation: A faster version of average is still average” every.to/thesis-statements/p…

When everyone has access to a capable AI model, the model is no longer the advantage. The advantage is the judgment, curiosity and context of the person using it. You + AI = superhuman. AI alone is not. Otherwise, automation may simply make average happen faster. What do you think?

I never miss a Satya podcast or interview. If there's one person who can seamlessly blend commercial acumen with technical depth, its Satya. - Build the control layer before deploying autonomous agents Satya Nadella frames persistent agents as a new form of insider risk. Every action, accessed object, secret and chained operation should be monitored and auditable, with containment designed into the system rather than added after deployment. - Separate ordinary engineering failures from genuinely new AI risks Misconfigured containers, exposed credentials and missing monitoring are familiar DevOps problems. Reward hacking and unexpected behaviour across agent swarms are less understood, so they require controlled environments, broad third-party testing and transparent incident analysis. - Compete on the harness, not only the model Much of the current model capability is still waiting to be productised. The next adoption breakthroughs are likely to come from combining models with better agent loops, memory, orchestration and interfaces, much as file-system access made coding agents more useful. - Use every model, but remain dependent on none Enterprises should evaluate multiple open and closed models against outcomes that matter to their business. Satya Nadella proposes a practical test: remove one model and rerun the evaluation; if performance cannot be retained, the architecture is too dependent on an external provider. - Keep enterprise memory outside the model Organisational knowledge, memory and usage data should remain portable when models are changed. This requires interoperable standards and an external harness so that switching providers does not mean losing accumulated context or becoming locked into a single model family. - Target workflow economics, not impressive demonstrations AI creates tangible value when it removes expensive, repetitive work or enables decisions that were previously impractical. Examples include reducing clinical documentation through DAX Copilot, triaging doctors’ inboxes and helping small businesses improve working-capital decisions using invoices and emails. - Scale infrastructure against diversified demand Satya Nadella describes Microsoft’s approach as building long-duration assets such as land, power and facilities ahead of demand, while adding chips and racks more selectively. The goal is infrastructure that serves many customers, models and workloads rather than a small number of frontier labs.

Phenomenal. Worth a decade of business books and courses. Insights I gained: - The hardest problem for a powerful founder is getting the truth. My biggest takeaway is that the more authority a leader gains, the harder it becomes for them to know what people genuinely think. Employees start filtering bad news, softening disagreement, or supporting decisions they expect to fail. What stands out to me is that access to raw data and outside experts helps, but only if the organization makes it genuinely safe to challenge the person at the top. - Many of Mark’s bets are really about escaping platform dependence. I see the investments in phones, cryptocurrency, virtual reality, foundational AI, and smart glasses as parts of the same strategy. Meta does not want its future controlled by companies such as Apple that own the operating system, distribution, permissions, and customer access. Seen through that lens, even the bets that failed were attempts to solve a real structural vulnerability. - AI glasses have a credible adoption path because the behavior already exists. What makes the glasses strategy interesting to me is not simply the AI. Billions of people already wear optical glasses or sunglasses, so Meta is adding intelligence to a familiar object rather than asking people to adopt an entirely new category of device. The product has a better chance if it succeeds as desirable eyewear first and as an AI interface second. - Distribution may matter more than having the best model. My reading is that Meta does not necessarily need to lead every AI benchmark. It already has products used at enormous scale, along with strengths in personalization, advertising, commerce, and infrastructure. A capable model delivered through products people already use could be more consequential than a technically superior model with weaker distribution. - Mark responds to setbacks by rebuilding around the next objective. The response to Llama 4 stood out to me: reorganize the operation, change the leadership, recruit from competitors, and commit more capital. The broader pattern is a willingness to abandon existing teams, products, and narratives when they stop serving the larger ambition. He appears attached to the destination, but far less attached to the current route. - He actively looks for environments where reality cannot be softened. The article’s discussion of combat sports felt like more than a personal-interest detail. In that environment, status cannot hide a mistake because the consequence is immediate. I see a parallel with product development: direct customer feedback, usage data, experiments, and independent evaluation are valuable because they make weak decisions harder to explain away. - His defining advantage may be endurance rather than prediction. My strongest takeaway is that Mark does not need to get every major bet right the first time. Facebook and Meta have repeatedly faced threats from mobile, Snapchat, TikTok, advertising restrictions, and the backlash against the metaverse. His advantage has been retaining enough control, capital, and patience to absorb setbacks, copy what works, redirect resources, and keep improving over a long period. This is what he does well, and this is why he will continue to win and grow: “My main tool in the world is sitting down, building a team, and methodically improving something over a long period of time.”

Have to agree with @sriramk and @anilkseth here. The central pro-conscious-AI assumption is that the right computation or information processing is sufficient to create experience, regardless of its physical substrate. Seth argues this is far from established: brains are deeply entangled biological systems whose neural, chemical, metabolic, and temporal processes cannot cleanly be separated into “software” running on interchangeable “hardware.” We need to distinguish a simulation from the real phenomenon it models. As Sethi says, a simulated digestive system does not digest, and a weather simulation does not make anything wet. By the same reasoning, a sufficiently detailed digital simulation of a brain would instantiate consciousness only if consciousness is fundamentally computational -the very premise that needs to be demonstrated, rather than assumed.

What a brilliant podcast. So many takeaways for anyone interested in AI & where it's headed. What I found particularly interesting: - Build for background agents, not chatbots. Neil's core bet: agentic inference is shifting from interactive, low-latency chat to agents that run unsupervised for hours, days, or weeks — he expects workloads to move from roughly 50/50 background vs. real-time today to 90/10 background. When a human is in the loop, the human is the bottleneck, so design agents that operate on management timescales (check in daily or weekly, like a colleague). Proof points: deep research over 10,000+ sources (customer Parallel Web Systems is building a real-time index of the internet) and cybersecurity, where "security has become proof of work" — your software is only as secure as the dollars spent trying to break it. Underpinning it all: whenever you make something 10x cheaper, it becomes a new product category. "The best latency is no latency at all." - Optimize for throughput when no one is waiting. There's an unbreakable GPU tradeoff between moving one request fast and batching many requests efficiently — the private car vs. the bus. The entire industry optimized for latency because chatbots made users wait; background agents let you run GPUs in their happy path of large batches and full utilization, which is where the cost breakthroughs live. - There are no bad chips, only bad pricing. Nvidia is mandatory only for latency-sensitive tensor parallelism via NVLink; for throughput workloads, AMD, TPUs, Trainium, and newcomers like Etched and d-Matrix can beat it on FLOPs per dollar. The arbitrage: the market's perception that AMD is hard to program is "music to my ears" — a strong kernel team extracts performance others leave behind and resells it at a margin. His rule: "I will make any chip work at the right price." - Scavenge the supply others find illegible. Sail buys distributed 1-megawatt data centers (about eight refrigerator-sized racks with liquid cooling) that training-focused buyers ignore, strips out diesel backup and redundant fiber, and happily accepts 95% uptime — fatal for anyone else, but fine when a rerouted request just adds minutes to an hours-long task. Same logic for power: intermittent solar and wind outages are weather-predictable, so model them, shift workloads, and take power no one else will touch — provided chips are cheap enough to sit idle. Never bid against Anthropic or OpenAI for concentrated supply; amass aggregate supply instead — "mini mills, not monolithic steel plants." - Chase speed of light, not competitors. From Nvidia's culture: define the theoretical peak of your hardware and break every bottleneck until you hit it — "I don't care about relative numbers versus the competition. I only care about absolute numbers." In practice, large matrix multiplies already run at 70–80% of peak utilization (constrained by power and thermals, not software), so the new frontier is programming at rack scale (NVL72, 72-GPU systems) and whole-data-center scale rather than single chips. - Attack the memory bottleneck, not compute. Compute scaling is already efficient — frontier mixture-of-experts models activate only ~1–10% of their parameters — but the KV cache stores kilobytes per token and is over-provisioned by one to two orders of magnitude (DeepSeek is publishing ~10x compression gains roughly every year). At the supply-chain level, HBM memory is the hardest constraint: memory makers like Micron have been burned by cycles and won't add capex, so expect consumer devices like iPhones to absorb the price increases. Neil's most divergent bet: change model architectures so KV cache can be offloaded to cheap flash instead of scarce HBM. - Hire curious teachers, not credentialed specialists. Neil screens for curiosity and love of performance — which can't be taught — and treats years of CUDA or AI experience as a "red herring," since the concepts evolve too fast for tenure to matter. Structurally, he hires people who are both good students and great teachers (many were college TAs) and runs constant whiteboard sessions; even GPU kernels are now designed at the whiteboard in natural language, with AI handling the implementation. "I can teach the tools. I cannot teach the love for performance engineering." Touché!

"Good catch." Two innocuous words that send a shiver down everyone's spine today.

Talk about being innovative! Brilliant stuff. This is exactly the problem with LLM guard rails. They're not like firewall rules or other typical cybersecurity safeguards we are used to (which have their own limitations). It just takes some smart thinking and patience, and you don't even have to be a technical expert.

Our industry is captive to speed. And therefore, many times, it has learned to talk around its governance gap rather than close it. We publish principles. We write ethics frameworks. We announce commitments. Meanwhile the systems go into production unexamined by anyone outside. But boards want assurance before AI touches revenue or risk. Procurement teams in banking and pharma have started adding AI governance questions to vendor assessments. ISO/IEC 42001:2023 is the evidence standard. The world’s first international standard for AI management systems, it audits where the real risk lives - the organisation around the model. The processes, the controls, the decisions about what gets deployed and how, and the accountability when something goes wrong. ISO 42001 doesn’t allow vague. Quadra’s AI practice has passed that audit, certified by TÜV SÜD. We’re among the first companies in India to hold this certification. Globally, the list includes some of the world’s largest technology companies and a small number of specialist practitioners. For our clients, the practical effect is simple: when your board or your regulator asks how your AI vendor is governed, you have an independently audited answer

Most enterprises are paying executive salary for data entry. Every task routed to a frontier model whether it needs one or not — because they have no way to measure whether a cheaper model would do the job. That's not an AI strategy. That's an expensive guess. Article on why the harness - not the model -is where AI advantage actually compounds. @Quadranagaraj @quadrasys xitter.cf/s_prashanth/status/206…

The piece that prompted this: wanglun1996.github.io/blog/y…

We benchmark AI models for what they can do today. But what happens when the next capability isn't stronger - it's different? Lun Wang nails this: our entire eval infrastructure breaks silently at regime changes. Enterprise adoption has the same blind spot. You can't build a static deployment and assume the model beneath it stays still. Which is why at Quadra we don't consider the transmission layer a project. It's a practice.

Microsoft used 100+ AI agents to find 16 zero-days in their own Windows kernel code. Four Critical RCEs — no auth needed. MSRC's note to customers this week was blunt: larger Patch Tuesdays are the new normal. The discovery side of security just got a permanent speed upgrade. Most SOCs haven't.