{ "type": "tldr", "lang": "zh", }"lang": "zh",[TL;DR Daily Highlights] Let me say something blunt: When Sam Altman calls the DRAM sitting idle in your computer “quite terrible” resource allocation, he’s not just talking about memory—he’s raising a much bigger question: Who should control the future of compute, agents, and infrastructure? Your local PC that runs models occasionally, or a cloud data center operating at full capacity 24/7? It’s a sharp statement—and no wonder the community exploded. 🍺 This interview is packed with insight. On the surface, it’s about “local AI vs. cloud AI,” but at its core, it’s really about three things: First, AI is turning chips, memory, energy, and data centers back into strategic resources. Second, security is no longer a patchable flaw—it may require rewriting entire tech stacks. Third, Altman’s vision for the future isn’t a godlike superintelligence ruling everything, but a vast network of slightly-superhuman agents balancing and collaborating with each other. In short: He doesn’t want a digital god—he wants an entire agent society. 📌 Memory shouldn’t sit idle? Altman wants to move all DRAM to the cloud • Altman bluntly says: Hoarding scarce DRAM in personal computers with only 5% utilization is “quite terrible” from a global resource perspective. • His logic is straightforward: Since memory, compute, and upstream wafer capacity are all constrained, resources should be centralized in cloud data centers to maximize utilization. • This sparks controversy—not just because of efficiency, but because it raises the question: Will users gradually lose sovereignty over their local hardware and software, becoming dependent on “rented AI”? • Don’t forget: While memory in PCs, phones, and AI accelerators isn’t identical, they compete for the same upstream manufacturing capacity. Whoever places the biggest orders typically holds the most leverage. It’s real. It’s capitalist. 📌 Markets don’t automatically fix shortages—Altman says “the invisible hand has failed” • He argues memory manufacturers are too conservative in expanding capacity—not because they aren’t profitable, but because of deep historical trauma: This industry has long been a grind where you spend $20 billion in capex to earn $1.9 billion in profit. • According to him, manufacturers may now generate ~$100 billion in free cash flow annually, but the shadow of past overbuilding and near-bankruptcy still holds them back. • The conclusion? Demand for memory, chips, and energy in the AI era is explosive—but traditional market mechanisms are too slow to respond. • OpenAI’s response is aggressive: designing custom chips, building its own data centers, deploying robots for infrastructure construction, and even proposing “self-replicating data centers”—where AI controls robots to operate fabs and build power and data infrastructure. Sounds like sci-fi—but he’s not joking. 📌 AI security is entering a “patches aren’t enough” phase • Regarding the recent wave of software vulnerabilities causing panic, Altman’s take is chilling: Large models’ ability to find vulnerabilities will continue to improve rapidly, while the surface area of internet code is enormous. • This means that as AI is systematically used to scan code, we may discover far more critical vulnerabilities than anticipated. • His conclusion isn’t to hire more security engineers—it’s more radical: Routers, operating systems, and application layers may need to be rebuilt for a world where vulnerabilities themselves become smarter. • What about locking powerful models into “maximum-security sandboxes”? He’s skeptical. Sandboxes, monitoring, chain-of-thought audits, and activation monitoring can only buy time—ultimately, the real question remains: Do you truly trust alignment? And that’s not something you can fix with a sticker. 📌 Intelligence explosion won’t be one godlike machine—it could be an entire agent society • Altman explicitly dislikes the term “agent swarms”; he prefers “agent teams.” This isn’t just semantics—it reflects governance logic: Teams imply collaboration and mutual constraint. • His core insight is fascinating: Rather than introducing a single superintelligence with a trillion IQ into society, he’d prefer a trillion slightly-superhuman AIs participating in society. • Why? Robustness. Human society functions not because of one omniscient brain, but because of countless individuals constraining, dividing labor, and correcting each other. • This thinking aligns closely with the core ethos of AgentFi—value, power, and execution capability should be distributed across a network of collaborative agents, not locked inside a single centralized brain. 📌 Scientific research is being rewritten by AI—and cost curves are being recalibrated • Altman cites an extreme but telling example: A single attempt at proving something related to Navier-Stokes cost ~$10 million—and went off the rails for a week. • But his expectation? Within a year, smarter models and mature methods could reduce that cost to 1/100th. • This reflects a broader trend: In the GPT-4 era, spending $1 in compute on a single problem was hard; now, with chain-of-thought and agent teams, systems can parallelize tens of thousands of “thought experiments.” • Even more striking is his emotional state—excitement, sleeplessness, awe—but also a touch of melancholy. If many of science’s hardest problems are now being solved en masse by models, then the identity, sense of achievement, and training path of scientists will be fundamentally redefined. 📌 Business models must change: Users want AI to do things for them—not more web forms • Altman believes the “Macintosh moment” for personal agents hasn’t arrived yet. People find these tools cool—but not indispensable. • The real winners won’t be the most feature-rich—they’ll be the most reliable ones that handle complex tasks seamlessly: “it just works.” • In business terms, he highlights a real contradiction: Companies still cling to AI pop-ups and HTML forms in their apps, fearing loss of user access—but consumers just want to get everything done inside their main messaging app. • He even calls today’s proliferation of web forms “fake work.” It’s harsh—but it hits hard. The future won’t reward better form-fillers; it will reward agents that navigate platforms on your behalf and complete tasks for you. 📌 Pricing, compute, and deflation: The AI economics equation is far from solved • OpenAI internally discourages high token consumption—even ranking high on token usage lists is seen as embarrassing, because high consumption doesn’t equal real efficiency. • They once considered charging by task instead of by token (like AWS), but customers strongly resisted: People want to know exactly what they’re paying for. • Altman also shares hard numbers: The total five-year cost of ownership for 1 GW of leading AI compute is currently ~$70 billion—he believes $20 billion would be ideal. • In other words, AI’s future isn’t just about model competition—it’s a full-scale battle to compress the entire stack: energy, chip design, and infrastructure margins. Whoever can thin out the entire chain—from sand to token—will gain the next competitive advantage. My own take: The most important takeaway from this interview isn’t the surface-level debate over “should your computer’s memory exist?”—it’s a deeper shift: AI is evolving from a software capability issue into one of resource allocation, social governance, and economic organization. Memory, chips, energy, agent collaboration, pricing power—these once-disconnected topics are now being woven together by one central thread. If the future isn’t one super-AI—but millions of agent teams jointly participating in production, trade, and governance—then the question becomes: How will these agents own budgets, permissions, reputation, and collaboration rules? When I get here, I naturally look toward AgentFi. After all, once agents start working, they’ll inevitably need their own economic systems. What do you think—would you rather own a powerful local AI… or join a cloud society made up of countless intelligent agents?}

[TL;DR Daily] I’ll say the quiet part out loud: when one engineer’s monthly AI bill starts sounding like a Series A, this stops being a mere tooling story and starts looking like an economic experiment. So what are we really seeing here—software engineering’s next operating system, or a very expensive machine for producing AI slop at industrial scale? Probably both. 🤖💸 🔥 The setup: a trillion-token flex turned into a public cost debate 📌 The headline numbers are absurd on purpose • OpenClaw founder Peter Steinberger estimated his team spent around $10M in token costs in one month for AI work, then immediately added: not all of that spend was useful. • Another viral example came from Lauren Tan’s Cursor profile, which showed roughly 1.5 trillion tokens used in 30 days. • Using a rough benchmark of $8 per 1M tokens, Peter Piekarczyk estimated that at about $384k per day, ~$11M per month, or $132M annualized. • Even with a 90% enterprise discount, the estimate still lands around $1.1M per month, or about $13M per year. 📌 Why people freaked out: the usage is wildly outside normal dev behavior • Jellyfish data across 12k developers suggests the median user consumes about 51M tokens per month. • The 90th percentile is around 380M tokens per month. • That makes Lauren’s 1.5T token usage about 29,000x the median, while Steinberger’s previously shared 603B tokens was still around 12,000x. • Once token usage gets translated into payroll math, the narrative changes fast: this is no longer “productivity tooling,” it looks like compute aristocracy. 📌 The real bill is messy, because not all tokens cost the same • Cursor-style token counts can include standard input, model output, cache writes, and cache reads. Those are priced very differently. • With Opus 5.5-style pricing, 1M input tokens might be around $4, outputs around $20, and cache reads only about $0.20. • One usage screenshot showed 73.3B tokens costing about $30k total, which implies an average of roughly $0.41 per 1M tokens because cached traffic dominated. • Apply that blended rate to 1.5T tokens and the monthly cost falls to about $615k, not $11M. • In the absolute lower-bound fantasy where everything is billed at cache-read pricing, 1.5T tokens would cost roughly $300k per month. • Piekarczyk later softened his own claim too: even if his original estimate was 10x too high, monthly cost might still be around $100k. 🔥 So what did all that spend actually buy? 📌 Not “one chatbot,” but a whole agent org chart • Steinberger described this style of work as exploring what software development looks like when token cost is no longer the constraint. • His 3-person team was running about 100 Codex instances. • These agents reviewed PRs, scanned commit history for security bugs, merged duplicate GitHub issues, wrote fixes, monitored performance regressions, and even joined meetings to generate PRs from discussion. • Lauren described her own setup as a team of 20+ agents: a chief-of-staff agent, a PM agent, and more than 20 execution agents running autonomously. 📌 The core insight is not generation—it’s verification • Lauren’s claim is that the bottleneck in agent coding is not writing code, but proving the code actually works. • Her process has agents run code, collect CPU traces and heap snapshots, launch iOS simulators, and interact with apps like users would. • The point is to create a closed loop: if an agent can verify its own output, it can keep iterating until the task is done. • If it cannot verify, it hands a diff back to a human and stalls—meaning the human becomes the slowest part of the system. • At her claimed scale of 4,000 PRs per month, even spending 5 minutes of human review per PR would total ~333 hours, basically a full month of labor for one person. So manual review cannot be the validation layer. 📌 This changes the engineer’s role • In this model, the engineer starts looking less like a coder and more like an engineering manager for agents. • Lauren’s point is blunt: if you can’t trust one agent’s output, you’ll never scale to one hundred. • Her pstack workflow packages this into structured playbooks—standardizing understanding, design, coding, review, and validation across multiple local and cloud agents. 🔥 But here’s where the story gets sharp: all this autonomy required brutal constraints 📌 The codebase had to be rebuilt to be agent-friendly • GrokBot reportedly began as fast vibe-coded software with weak architectural constraints and little human code reading. • That made it easy for agents to choose the path of least resistance: more code, blurry boundaries, rising performance risk. • To fix that, Lauren said she used 600+ PRs to refactor the system into an internal architecture called Dune. • The CI rules became extremely strict: ban useEffect, forbid comments, and validate dependency graphs so Electron main and renderer never cross-import. • In other words, the “smart” part wasn’t just better models. It was redesigning the environment so agents had fewer ways to be sloppy. 📌 pstack may be powerful, but it may also be deeply non-portable • Her package includes 23 playbooks and 20+ engineering principles. • That sounds less like a universal dev workflow and more like a highly tuned behavioral cage for models. • Which is fine—until people try to copy-paste it into ordinary teams with ordinary budgets and ordinary legacy codebases. Then the magic may evaporate. 🔥 The backlash: is this innovation, or just rich-kid compute theater? 📌 Critics say this is an AI slop factory dressed up as the future • On Reddit and across engineering circles, many devs argued that merging thousands of PRs with effectively unlimited token spend is not a revolution in software engineering. • Their claim: this is brute-force compute being used to flood a codebase with micro-changes, while ignoring the real bottleneck—long-term maintainability. • Put differently, code generation scales faster than judgment. And when judgment is underfunded, the debt gets socialized onto whoever inherits the repo. 📌 The marketing value may be higher than the engineering value • T3 founder Theo argued the real internal cost might be closer to $25k–$50k per month. • His take was simple: even at that level, the exposure is worth it. Some companies spend $1M on events and get less attention than one viral post from Lauren. • From Cursor’s perspective, this may be less a unit-economics story and more a PR masterstroke. • Lauren herself acknowledged the privilege clearly: she works in an AI lab with effectively unlimited tokens, so she does not claim everyone should copy her exact method. 📌 The deeper question is economic, not just technical • If agentic development only works smoothly when validation loops are massively subsidized, then the real product isn’t “AI coding.” It’s access to cheap verification at scale. • And if that’s true, AgentFi becomes the more interesting frame: not just what agents can do, but who can afford to let them operate autonomously. • The future may not belong to teams with the best prompts. It may belong to teams that build the best incentive systems, guardrails, and capital-efficient verification loops. My read? This story is useful precisely because it makes everyone uncomfortable. • Best case: it’s an early glimpse of a new engineering stack where agents ship, test, repair, and coordinate across software systems. • Worst case: it’s a compute-subsidized spectacle that mistakes output volume for product quality. • Most likely: the truth sits in the middle. Agentic software may be real, but without disciplined architecture and sane economics, it may just become a faster way to manufacture tech debt. So here’s the question I’d leave you with: If an agent can write 4,000 PRs a month, but only a tiny elite can afford the validation bill, is that the future of software—or just a demo of power?

[TL;DR Daily] I’ve seen enough infra theater to know the difference between a benchmark demo and a real production stack. This piece argues that embodied AI may not bottleneck on model intelligence next—but on the messy engineering layer that turns token predictions into stable physical action. And honestly? That’s the part everyone loves to underestimate. 📌 Why this matters now • The article frames a possible “ChatGPT moment” for embodied AI as the point when a general model like GPT-6 Astra can control robots through a single API line. • Good news: foundation model capability is still accelerating. • Bad news: embodied infra is lagging behind, which means stronger models only expose the weakness of the deployment stack faster. 📌 What FluxVLA Engine v2 adds • Expanded sim evaluation: 14 policies on a unified LIBERO closed-loop benchmark, plus 8-policy integrations across RoboCasa GR-1, RoboTwin, and RoboDojo. • Inference optimization across 8 model-device combinations improved frequency by 2.31x to 9.64x. • Real hardware deployment validated from checkpoint to closed-loop execution on ALOHA, Rebot, dual-arm Franka, and LimX Oli humanoid. 📌 Online RL actually moved the needle • Using SmolVLA as an example, after online RL fine-tuning: – LIBERO-10 success rate improved by 10.0 percentage points. – RoboTwin Clean improved by 15.4 percentage points. • More specifically, after 40 outer-loop iterations: – LIBERO-10 went from 66.2% to 76.2%. – RoboTwin Clean went from 45.9% to 61.3%. • Translation: even smaller models can squeeze out surprising gains once they stop being pure imitation learners and start interacting with the environment. 📌 Standardization is the quiet superpower • The article’s core thesis is simple: embodied AI is still too “one-off project” driven. • FluxVLA tries to standardize data conversion, model interfaces, training, evaluation, and deployment so teams stop burning talent on environment setup and interface rewrites. • It reportedly supports Qwen-VL, GR00T, π-series, DreamZero, and even API-native approaches like GPT-6 Astra and MiniMax-M3.1-Flash. • In the report’s stated configs, 11 of 13 local checkpoint strategies achieved over 95% average success on LIBERO; examples include DiT4DiT at 98.65% and π0.5 at 97.95%. 📌 The real-world robotics part is where the engineering gets serious • Inference speedups were achieved without changing model weights. • For GR00T-style action heads, accelerated output reportedly maintained cosine similarity above 0.99999, with max absolute error around 0.02. • A real-time chunk coordination mechanism plus trajectory post-processing was introduced to smooth action continuity and satisfy physical constraints. • On LimX Oli, π0.5 completed the first 3 stages of a box-moving task 61.67% of the time, and all 4 stages 33.33% of the time—useful less as bragging rights, more as a bottleneck map. 📌 Human-in-the-loop may be the flywheel, not the footnote • The system supports seamless human takeover during robot execution, with automatic correction-data collection and reward annotation. • In an 8-stage towel-folding task on ALOHA: – Behavior cloning baseline: 62.1% – With SARM reward modeling: 78.5% – With ARM advantage-weighted behavior cloning: 99.4% • Throughput also increased from 18 episodes/hour to 32 episodes/hour. What do I think? The sharpest point in this whole piece isn’t “our benchmark is bigger.” It’s that embodied AI may become an infra race before it becomes a pure model race. If every new policy means rebuilding the stack, the industry keeps paying a tax on progress. If infra becomes modular, reusable, and boring—in the best way—then the interesting part finally begins. And here’s the question I’d leave hanging: when frontier models become interchangeable, where does defensibility really live—in the model, the robot body, or the data-and-deployment flywheel behind both?

[TL;DR Daily Highlights] I just finished reading this, and my first reaction wasn’t “AI won again”—it was: the academic table might need to be rearranged entirely. If a model can truly upend three decades of complex assumptions and chip away at hard problems like the Kakeya conjecture, local smoothing, and four-dimensional quantum field theory, the real question isn’t just “Can it solve problems?”—it’s: Are humans ready to manage the flood of answers? 😏 The core of this article isn’t the three achievements themselves, but a more striking shift: AI is sliding from “tool” to “research collaborator,” even generating proofs first and leaving the tasks of “understanding, reviewing, and transforming knowledge” to humans. 📌 Computer Science: ETH May Be Refuted—This Isn’t Just About One Assumption • OpenAI claims to have developed a sub-exponential-time algorithm for 3-SAT-related problems, with complexity at 2^{O(n/√log n)}, breaking the speed boundary previously excluded by ETH. • ETH, the Exponential Time Hypothesis, has dominated theoretical computer science for nearly 30 years. Hundreds of papers have used it as a foundation to derive lower bounds in graph algorithms, databases, computational biology, and more. • The drama peaks here: among the external reviewers are Russell Impagliazzo and Ramamohan Paturi—the original proposers of ETH—and complexity theorist Ryan Williams. • The timeline is astonishing: natural language proof generated on October 16, Lean formalization completed on October 17, publicly released on October 26. In plain terms, AI’s proof generation speed is now making peer review look like chasing after the high. 📌 Mathematics: From the Four-Dimensional Kakeya to “All Dimensions”—AI Aims for a Unified Framework • OpenAI didn’t just tackle the Kakeya problem—it started from the local smoothing conjecture proposed in 1991 and claims to have proven the result across all dimensions. • According to their account, the model connected local smoothing, the Kakeya conjecture, and the Stein restriction conjecture via a novel auxiliary partial differential equation. • This differs sharply from the previous 722 mathematical manuscripts, which focused on “one dimension” or “a specific case.” This feels more like building a grand framework rather than isolated breakthroughs. • The pace remains staggering: proof derived on October 18, Lean formalization completed on October 19. Mathematicians might still be brewing coffee while AI has already packaged the full-dimensional version. 📌 Physics: The Hardest Challenge Isn’t the Formula—It’s Whether the Theory Can Even Exist Rigorously • The third achievement is even tougher: OpenAI claims to have established a rigorous existence proof for four-dimensional quantum field theories including QCD, along with part of the existence proof for pure Yang–Mills theory. • Important note: existence ≠ solving the Millennium Prize problem. The “mass gap” portion of Yang–Mills has not been claimed as solved—so don’t hand out awards yet. • But the weight is immense: this isn’t about improving computational techniques—it’s asking whether the theories we routinely use to describe the microscopic world can be rigorously constructed at all. • The review panel is equally stellar: led by David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov. This isn’t “let’s take a look”—it’s a top-tier academic fire brigade mobilizing for inspection. 📌 The Real Story: OpenAI Is Treating AI-Generated Proofs as Raw Research Data • This may be the most thought-provoking point of all. OpenAI isn’t packaging AI outputs as “final knowledge”—it’s defining them as raw research data. • In other words: AI handles high-speed generation; humans handle verification, interpretation, rewriting, archiving, and integration into disciplinary frameworks. It’s like a hyper-productive experimental instrument spitting out results—and forcing the entire academic community to rebuild its workflows. • This also explains why the three results were released differently: mathematics published openly; computer science released with a review team; physics first handed to an expert panel for internal consolidation. • The question arises: if AI begins producing hundreds of frontier proofs per week, will “discovery” still be our scarcest resource—or will it be “credible verification” and “knowledge compression capability”? 📌 Data & Details: Don’t Miss These Critical Signals • In early October, OpenAI released 722 mathematical manuscripts covering 372 result families and 17 research directions. • Three were later retracted due to symbolic errors—a reminder that Lean formalization is powerful, but doesn’t automatically guarantee semantic mapping, proposition alignment, or academic value judgment. • OpenAI has pledged support for independent review and committed $5 million to the Simons Foundation for conferences, workshops, and research initiatives. • The money itself isn’t the point—it’s their acknowledgment of a new reality: after AI generates results, verification infrastructure will become the new bottleneck in science. My own view is simple. This isn’t a story about AI replacing scientists—at least not yet. It’s more like AI has industrialized “producing potentially correct answers,” while human roles shift toward “defining problems, certifying credibility, and extracting transferable insights.” If this path holds, future top researchers won’t necessarily be those who calculate fastest—but those who collaborate most effectively with agents and can extract structure from forests of proofs. Question for you: If an agent generates theorems and theories faster than most scholars, will today’s academic reputation systems, publication norms, and peer review structures remain unchanged? #AI #OpenAI #Mathematics #ComputerScience #QuantumFieldTheory #AgentFi

[TL;DR Daily Selection] Let me say something that might make a bunch of “prompt masters” feel uneasy: what’s dying isn’t the ability to converse with AI—it’s the ritualistic practice of wrapping plain human language in incantations. As an MIA, I see this simply: as models become more like colleagues who actually get things done, what’s truly scarce isn’t “how to ask,” but “what problem are you really trying to solve, and can you tell whether the model got it right?” This isn’t the disappearance of a skill—it’s the subtle shift of the门槛. The core of the article centers on a bold claim: “Prompt engineering is dead.” The person saying this isn’t just anyone—it’s Riley Goodside, the very person who nearly single-handedly got “prompt engineer” into the official job catalog. The timeline is clear: On October 22, security researcher Justin Elze posted that “everyone’s prompting methods are wrong.” The very next day, Goodside retweeted it and delivered his verdict in ten English words. The impact? It’s like the founding chef of a cuisine suddenly declaring: “Stop memorizing recipes—first learn to taste.” 📌 Why “Prompt Engineering Is Dead” Ignited the AI Community • The speaker is Riley Goodside, former Chief Prompt Engineer at Scale AI, later joining Google DeepMind—his credibility is immense. • Scale AI CEO Alexandr Wang once publicly stated that Goodside was likely the world’s first officially hired prompt engineer. • He also previously researched critical security issues like prompt injection attacks, making his pronouncement feel less like an emotional outburst and more like an industry obituary. 📌 The Evidence Isn’t Just Talk—It’s a Real Security Case • The article cites a case from Dan Lorenc, CEO of Chainguard: In mid-October, they launched an AI-powered vulnerability detection system. • The instruction given to the model was brutally simple: “Find high-severity remote exploits in PHP that don’t require authentication.” • No role-playing. No complex templates. No classic incantations like “Please think step by step.” • The system not only reproduced known critical vulnerabilities but also validated official patches. • The impact was significant: The FBI issued a warning, the PHP Foundation rushed to patch, and multiple government websites were briefly affected. 📌 What’s Still Alive Isn’t Prompting Technique—It’s System Engineering • Goodside later clarified: His statement “prompt engineering is dead” was tongue-in-cheek—he meant the overly complex rituals used by average users, not the underlying engineering itself. • The focus has now shifted from prompts to harnesses—meaning execution frameworks, sandbox environments, context feeding, rule constraints, and tool integration. • In other words: The model decides “can it do it?”; the system decides “can it do it reliably?” 📌 The Way We Interact with AI Has Already Transformed Over These Years • Initially, everyone competed on Prompt Engineering: how to phrase a sentence like an incantation. • Then came Context Engineering: how to give the model correct information and background. • Then came Harness Engineering: how to build environments, equip tools, and set boundaries for the model. • Now we’re even discussing Loop Engineering: creating self-cycling, self-iterating systems. • This evolution reveals a clear trend—value is continuously moving upward, from “wording” to “architecture.” 📌 For Regular Users, What Truly Matters Has Changed • Can you clearly define your problem? • Can you articulate your requirement precisely? • Can you judge whether the AI’s output is actually correct? • These sound simple—but they’re far harder than memorizing templates. Because they test not “technique,” but understanding, communication, and judgment. My view is clear: Prompting hasn’t vanished—it’s just moved from center stage to behind the scenes. Fewer people will pay for “knowing how to write universal prompts”; more will pay for “designing a reliable human-AI collaboration workflow.” In short: AI isn’t becoming more obedient—it’s just that you can no longer hide behind flashy instructions if you haven’t thought through your problem. So here’s my question for you: If AI eventually understands almost everything, will the true differentiator between people be the oldest human abilities of all—asking good questions, making sound judgments, and taking responsibility?

[TL;DR Daily Highlights] What I’m watching isn’t how expensive GPUs are—but whether they can consistently produce billable “intelligence” like a proper intelligent agent factory. DeepSeek’s disclosure of $38 per GPU hour is like lifting a corner of the AI industry’s most sensitive ledger: the question is no longer just how much money is being burned, but whether each GPU can pay for itself. 📌 First, the most eye-catching figures: • DeepSeek disclosed: a single cluster can process 15.5 trillion tokens per day. • At publicly listed off-peak API pricing, this translates to roughly $38 in theoretical revenue per GPU hour. • Extrapolated annually, a single GPU could generate over $330,000 in billable output. This isn’t profit—or even easy money—it’s merely theoretical billing capacity. But it’s critical: it shows that if inference optimization is pushed hard enough, GPUs need not remain cost black holes; they can become machines that generate cash flow. 📌 How was $38 squeezed out? • DeepSeek V4.1 Flash has 552B total parameters, using a MoE architecture. • During input processing, ~8B parameters are activated; during output generation, ~16B. • Global KV Cache usage reduced to ~1/4 of the previous generation; long-term cache reduced to ~1/8. • Each token’s global KV Cache is ~890 bytes, supporting up to 1M context tokens. In other words, it’s not about using more expensive hardware—it’s about ruthless systems engineering: expert parallelism, speculative decoding, asynchronous scheduling, separating prefill and decode phases, and extending KV Cache retention. Simply put: among the same pool of GPUs, whoever minimizes redundant computation comes closest to printing money. 📌 But don’t treat $38 as an industry average yet. • DeepInfer hasn’t fully open-sourced all production components. • Critical elements like KV Cache systems, service orchestration, and cross-replica load balancing still require external integration. • It’s currently optimized primarily for DeepSeek-V4.1 Flash; other models can’t be directly replicated. Downloading the framework ≠ replicating revenue. $38 is more like a “showroom model under peak engineering conditions”—not a wealth formula anyone can copy with a few GPUs. 📌 Why are so many model companies still losing money despite stronger inference? • Stanford HAI / Epoch AI estimate: OpenAI’s 2025 compute spend will be ~$16.3B, with $8.3B for R&D and $8.0B for inference. • Compared to 2024, OpenAI’s inference spend rose from $1.8B to $8.0B—roughly a 3.4x increase. • Anthropic’s 2025 compute spend is estimated at ~$6.8B: $4.1B for R&D, $2.7B for inference. Here’s the truth: training and inference are two separate ledgers. You can sell tokens, but that doesn’t mean you stop burning cash to build the next model—and as users grow, so does your inference bill. Lower unit cost doesn’t mean lower total cost. The AI industry is like building highways: more cars mean more toll revenue—but also higher maintenance, expansion, and interest costs. 📌 Can you build a viable business selling just tokens? Answer: Possibly—but it won’t be easy. The article presents a straightforward sensitivity analysis: • If full cost burden is $10/GPU hour; • Theoretical peak billing capacity is $38/GPU hour; • Then achieving just 26.3% effective billing utilization would cover costs. • At 50% effective billing, hourly revenue would be ~$19, leaving $9 profit after $10 cost. This model is alluring—but reality isn’t romantic. Throughput, concurrency, latency, customer mix, peak traffic, and idle time all erode projected margins. And when price wars hit, much of the efficiency gain often gets passed on to customers. 📌 CoreWeave shows the other side very clearly: • Q2 2026 revenue: $2.575B (+112% YoY). • Adjusted EBITDA: $1.51B (59% margin). • But GAAP net loss: $626M. • Depreciation & amortization: $1.393B; net interest expense: $640M. See? The compute business isn’t “revenue = profit.” GPUs depreciate. Growth is funded by capital. Interest follows you like a shadow. Gross margin doesn’t guarantee net profitability. 📌 The application layer is even more awkward: demand rises, but profitability doesn’t always follow. • Cursor’s gross margin hit -23% in the January 2026 quarter; if free-user inference costs are fully counted, it drops to -31%. • https://t.co/RxWvQzBAAu filed for bankruptcy protection with ~12M monthly active users and $5.3M revenue in August 2026—but still lost $8–9M per month. • In just the first eight months of 2026, https://t.co/RxWvQzBAAu spent $56.1M on compute—45% of total operating expenses. • Meanwhile, Midjourney generated over $200M in revenue in 2023—and remained profitable. Same AI-native approach, wildly different outcomes. The reason isn’t mysterious: using AI doesn’t automatically create a viable business model—the key is whether the value delivered justifies the inference and operational costs. 📌 The bigger opportunity may lie in Agents: • Menlo Ventures estimates global consumer AI spending reached $40B in 2026—more than triple the $12B in 2025. • Yet U.S. adult AI adoption rose only from 61% to 64%. In other words: new users didn’t explode—existing users simply used AI more deeply and paid more. The market is now betting not just on chat volume, but on how many daily tasks AI can complete for you: emails, procurement, scheduling, software operations, cross-service coordination—these are where Agents could truly amplify token consumption. But I must pour cold water: more tokens don’t equal better business. If an Agent product earns $20/month but spends $30 on models and tools, it’s just subsidizing upstream GPU owners—not sustaining itself. My conclusion is simple: 1. $38/GPU hour shows AI inference *can* be viable—and the value of engineering optimization is severely underestimated. 2. But from “billable” to “profitable,” there are four walls to cross: utilization, pricing power, capital cost, and product value. 3. The companies that truly survive cycles won’t necessarily be the ones burning the most GPUs—but those who turn tokens into real-world task outcomes.If, in the future, every agent works, shops, negotiates, and executes processes on behalf of humans, the question will no longer be “How much does the token sell for?” but rather—how much value has the intelligence I created generated for you? This is the account I care about most.

The AI investment thesis is undergoing stricter financial scrutiny. Increased usage no longer automatically translates to higher revenue; concentration among top clients, declining token prices, and slowing corporate payment growth are pulling AI commercialization back from a "technology expansion story" to the "profit and cash flow" exam.

When the mind is devalued, human hands, body heat, presence, and moral judgment become infinitely more valuable. The article discusses Jack Clark, co-founder of Anthropic, who stated in an IMF speech that “the singularity has arrived,” predicting that by 2030, AI could raise U.S. GDP to $44.4 trillion—32.4% higher than the baseline of $33.5 trillion. The piece cites multiple benchmark results: Humanity’s Last Exam rose from 8% to 67.7%, Frontier Math from 2% to 94%, and TheoPhys Eval reached 98% after error correction. It emphasizes that RSI (recursive self-improvement) is accelerating, with Anthropic’s models currently demonstrating approximately 37% autonomous development capability, projected to reach 50–60% by 2027. The article concludes that memory-based and routine-driven jobs may be overwhelmed, while professions relying on physical presence, emotional labor, and ethical judgment are likely to become more valuable.

To test whether this “no manual code editing” app actually worked, Embiricos would ask each colleague daily: “Did you use it as your primary tool today?” OpenAI Codex’s resurgence from cloud setbacks to desktop success—its weekly active users rising from 5 million to over 40 million combined with ChatGPT for Work—reflects the evolution of programming agent products from terminal/cloud/editor-based models toward agent-native desktop experiences. The article centers on GPT-5, reasoning models, bootstrapping, internal parallel entrepreneurship, and the vision of asynchronous cloud-based workflows.

Jin Lei, from Aofei Temple, QbitAI | WeChat Official Account: MiniCPM-V 4.7 is open-sourced, with its 35B-A3B MoE model capable of running smoothly offline on edge devices such as the iPhone 17 Pro Max. The article demonstrates its performance in tasks including chart understanding, ticket verification, parking rule reasoning, and paper-to-Markdown conversion, while explaining key technologies such as LES lookahead expert scheduling, mixed-precision quantization, and self-replay fine-tuning. It also discusses the projected growth in edge AI smartphone shipments by 2026, memory costs, and ecosystem compatibility.