The Future of Artificial Intelligence and Proactive Personal AI Assistants

When we began comparing Claude Code, Codex, and other AI products, the conversation quickly turned to the harness. In products such as Claude Code and Codex, which provide a relatively complete working environment, the result depends on the model, its tools, and the workflow around it. Choosing a stronger model may do little for the quality of the work if the instructions are unclear, the environment is unsuitable, or the task has not been thought through.

My instinct is that Doubao Work or Doubao itself is already sufficient for a substantial range of everyday tasks. Most users may never push a model to its limits. More often, they may struggle to explain what they need, prepare the right information, or break a task into manageable parts. Differences between models still matter, of course, in complex coding, specialist reasoning, and work that requires repeated verification.

Claude had also been developing well before the attention generated by Claude Code. As I remember it, Claude was already attracting interest among users of tools such as Cursor before Claude Code became a broader topic of discussion. What later left a stronger impression was its ability to finish work within a complete environment. For ordinary users, that result is easier to appreciate than a discussion of model parameters alone.

One point carries into the rest of this discussion: as you spend more time using a model or product, it also shapes your habits. Past materials, tool configurations, and ways of working accumulate, allowing the next task to pick up where the previous one left off. This context that persists across tasks is an important part of how I understand an agent’s ability to attract and retain users.

ChatGPT notes · Models and harnesses jointly shape delivery

• Correction · A harness can organize context, tools, execution state, and feedback to help a model work more reliably, but it cannot eliminate differences in model capability. Whether a general-purpose assistant is “good enough” depends on the task. The sources cited here do not support a fixed percentage of ordinary users whose needs it can meet. ^[1][2]

• Explainer · Claude 3.5 Sonnet was released before the Claude Code research preview, establishing the sequence of model and product development. Public launch announcements do not establish when either became widely popular in a particular region, or that the harness was the main cause of that popularity. ^[3][4]

• Further thought · Doubao Work (豆包工作) is a specific product name. Feishu describes it as an AI work product natively integrated with Feishu; this should not be generalized into a merger of Feishu as a whole with Doubao. ^[34]

• Explainer · Context can include materials for the current task, conversation history, retrieved information, and tool state. It is related to usage habits, long-term memory, and switching costs, but acquisition and retention also depend on pricing, distribution, and the quality of the work delivered. ^[1][2]

When thinking about proactive personal assistants, I had been optimistic about products such as the “Doubao phone”: a combined hardware and software offering in which people speak to their phones and the built-in assistant carries out a range of actions. If the experience becomes smooth enough, it could change everyday habits and become an important route to monetizing AI.

My counterpart was more cautious. He pointed out that personal assistants have been finding their way into hardware for years. Companies including Huawei and Alibaba have already tried bringing voice or model capabilities to devices. A software company moving into hardware as AI develops still has to build industry experience, manufacturing capabilities, a supply chain, and distribution.

We also discussed Doubao’s Ola Friend earbuds. They give a voice assistant a wearable interface, but making that product possible leaves a set of commercial questions: why should someone buy another device? Does the added convenience justify the price? Can the cost, margin, and frequency of use support a lasting business? I find these harder questions than whether a model can be connected to the hardware.

There is also competition from established hardware manufacturers. Huawei has extensive experience in earbuds and smartphones, and some of its earbuds can already activate third-party applications such as Doubao. Users may not need to buy an AI company’s own device to access the same AI service. That makes differentiation even more important for a dedicated piece of hardware.

Combining hardware and software remains a path worth watching, but software expertise does not automatically become a hardware advantage. Industrial experience, manufacturing costs, delivery, and after-sales service all affect the final product. What I want to keep watching is whether that additional investment ultimately produces an experience that a phone app would struggle to offer.

ChatGPT notes · Hardware still has to justify its cost and frequency of use

• Explainer · Ola Friend is a real product: open-ear smart earbuds from Doubao that provide interaction through the Doubao app on a connected phone. Some Huawei earbuds can activate Doubao, subject to requirements involving the model, software version, and pairing setup; the feature should not be assumed to apply to every product. ^[5][6][26]

• Correction · The public information cited here does not establish Ola Friend’s overall sales, profitability, or commercial success or failure. Hardware costs and margins are valid commercial questions, but the product format alone does not show that a hardware strategy has failed. ^[5]

• Explainer · Voice assistants and smart hardware have developed over many years. Alibaba has also announced plans to integrate Tongyi Qianwen into products such as Tmall Genie. Responding to voice commands and continuously understanding a user’s needs well enough to act proactively require different capabilities. ^[7]

Returning to the application itself, we discussed three central questions for a proactive personal assistant: can it keep complicated interactions in the background, how much can it understand about its user, and can it move things forward on its own? Context led to the longest discussion. However capable an assistant is, it will struggle to become part of everyday life if it has to get to know you from scratch every time.

My counterpart was optimistic about platforms with large ecosystems, with WeChat and Apple as prominent examples. They are already part of frequent activities across communication, work, shopping, and entertainment, and can build continuous habits around them. I understand the reasoning: a platform closer to a user’s daily life has more opportunities to develop a comprehensive personal assistant.

Apple, for example, has built an ecosystem over many years through Apple Music, podcasts, communication tools, and the App Store. My idea is that if the operating system can connect applications through consistent interfaces, an assistant could obtain information and perform tasks more directly, without repeatedly moving between screens. As personal information and habits accumulate, switching to another platform could also become more costly for the user.

Alipay is another entry point worth watching. Payments, transport, and everyday services are closely connected to a person’s daily plans. Whether its partnerships with other platforms can extend further into shopping and work will influence how much an assistant can accomplish.

By comparison, a Doubao phone pursuing cross-application actions immediately encounters questions of interfaces and permissions. How much other platforms choose to expose, which actions can only be performed through the screen, and which behaviors are restricted all affect efficiency. Users want fewer steps; application platforms have business, security, and competitive interests to consider. These priorities will not always align.

A hardware partnership can bring an assistant into the operating system, but it cannot resolve every boundary between applications at once. A software company entering the smartphone ecosystem also has to work through these issues with manufacturers and developers. To me, this is both the appeal and the difficulty of a personal assistant: it promises one continuous experience while connecting services that each operate under their own rules.

The value of context ultimately depends on how deeply an assistant understands a person. My guess is that, with enough trust, users may tell their own AI assistant things they would not otherwise share openly. Information, preferences, and habits accumulated over time could make it increasingly personal and harder to replace. Platform ecosystems have an advantage here, but how much that advantage translates into useful capability depends on how well they use the information.

ChatGPT notes · Personal context depends on information quality and permissions

• Correction · Apple’s App Intents framework allows developers to expose actions and content that the system can use. Operating-system integration does not mean the system can compel unrestricted access to the internal data of every app. What developers expose, what users authorize, and the privacy design all remain relevant boundaries. ^[8][9]

• Correction · Ant Group, Alipay, and Alibaba’s Taobao should not be treated as one automatically interconnected data pool. An ecosystem of everyday services can provide entry points, but access to specific information and actions still depends on cooperation and authorization. ^[12]

• Explainer · The Doubao phone assistant already has operating-system-level partnerships with phone manufacturers. Screen interaction and tool calls can coexist, so its operation should not be reduced to Android accessibility features alone. Doubao’s published SAEP policy also defines a boundary for third-party apps to refuse screen automation. This is the platform’s own policy, not a uniform industry rule. ^[10][11]

• Further thought · Large ecosystems have advantages in access and use cases, but user-authorized connectors, open APIs, and on-device processing can also organize personal information. The value of context depends on relevance, accuracy, freshness, and retrievability. More information does not automatically produce better judgment, and users’ willingness to share private information needs to be examined in practice. ^[1][9]

Following that thought, what does a proactive personal agent actually mean? Traditionally, we first decide to do something, open an app, and enter a request. A future personal assistant that understands our work and life could notice what needs attention before we ask, and offer recommendations or make arrangements.

Suppose I discuss meeting a client next week. With authorized access to the relevant messages and calendar, the assistant could extract the agreement, check my availability, and prepare the schedule. I would not have to explain the same thing twice. An assistant that connects scattered information would be far more useful to me than another place to ask questions.

The idea can extend to travel, spending, and even financial planning. With knowledge of my itinerary, address, and account information I have authorized it to access, it could make plans better suited to my circumstances. Compared with a human personal assistant, it could continuously process more information without being limited to a few working hours each day. What it can see and how far it may decide on my behalf would, of course, have to be agreed separately.

In terms of interaction, I would like it to keep most of the process in the background and deliver the result directly. Users should not have to watch every search, file read, and tool call. The assistant should pause to discuss the parts that actually need a human decision. For the user, this brings a complex workflow into a single product experience: the work still happens, while the user no longer has to accompany the system through every step.

That raises another question: if the intermediate steps do not keep being rewritten as human-readable text, could some overhead be reduced? Once those repeated explanations are removed, could the system devote more of its computation to the problem that needs solving?

As this kind of assistant matures, it could take over some of the interactions now handled through basic apps. Opening a calendar, searching for travel details, and organizing messages could all become parts of a single task. As users rely on it over time, their habits and information would continue to accumulate. Commercially, the retention that comes from sustained use is an important part of the personal assistant’s value.

ChatGPT notes · Hiding the process and reducing computation are separate questions

• Correction · Turning a chat into a calendar entry is an authorized-use product scenario here. It should not be read as a claim that WeChat already reads every conversation continuously and schedules events by default. Its current terms describe use of relevant content in connection with user actions, instructions, and tasks. A tentative intention also needs to be distinguished from a confirmed arrangement. ^[13]

• Explainer · Reading account information, making recommendations, and executing transactions are separate capabilities with different permissions. Even with access to more data, an assistant must deal with errors in judgment, scheduling conflicts, authorization boundaries, and the consequences of its actions.

• Correction · Hiding intermediate steps does not automatically save computation. Reasoning that is not shown to the user may still consume tokens. Reducing unnecessary generation, compressing data transfer, and using structured data instead are separate optimizations that need to be evaluated. ^[19]

• Further thought · Even if an assistant takes over the user-facing interaction, services such as calendars, payments, and travel booking may continue to operate as tools or APIs behind it. Sustained use also depends on reliable delivery: users should be able to inspect and correct memories, understand consequential actions, and stop or take over a task when needed. ^[1][2]

The conversation eventually returned to automation. My counterpart suggested looking at AI along a line of development stretching back to the Industrial Revolution: machines gradually took over physical labor, computers assumed large amounts of information processing, and today’s agents are attempting more complex knowledge work. Seen this way, many apparently new products have something in common with earlier technological changes.

He was optimistic about the potential of AI companies such as Anthropic. Apple has shown how large a technology platform can become; companies built around models and information processing may be able to replicate their services even faster. It made me think that much of the potential attributed to AI companies comes from how much work once dependent on people they can turn into continuously operating services.

Underlying computing capabilities, including GPUs, are an important condition for that expansion. Tasks suited to parallel execution can run faster, and established methods can be deployed repeatedly in new settings. What interests me is the combination: as compute, software, and task organization improve, can a company serve more customers without increasing headcount and cost in the same proportion?

From steam engines and industrial production to software and agents, automation has kept expanding our expectations of efficiency. I want to follow that thread further. We are currently trying to get AI to do work; the next step may also change how information moves—how it is found, judged, and processed, then passed to the next part of a system that can act on it.

If these processes can be replicated consistently, larger companies with a wider reach may indeed emerge. Still, it is difficult to say in advance exactly what “information automation” will look like. I think its scope will depend on whether complex tasks can actually be broken down into work that can be delivered reliably.

ChatGPT notes · Automation offers one lens on industry change

• Correction · Automation is a useful lens for technological change since the Industrial Revolution. It cannot account for all of human history or explain company scale and market value on its own. Claims about information distribution and business expansion also need to consider demand, costs, competition, and institutions.

• Explainer · Anthropic is the company; Claude is a model and product brand. Optimism about its potential is a judgment, not evidence of an IPO timetable, valuation, or investment return. No such forecasts are added here. ^[25]

• Correction · There is no fixed numerical conversion between GPUs and CPUs. Actual acceleration depends on factors such as task parallelism, algorithms, numerical precision, memory, and communication. Hardware speed also cannot be converted directly into a company’s rate of growth. ^[14]

When discussing the development of AI itself, I initially tried to connect the Transformer, the agent, and the harness along a single line: the model processes information, agents divide up the work, and the harness organizes execution. Each can help automate work, but I had not fully understood how they connect. In particular, whether a model’s attention mechanism has a meaningful counterpart in collaboration between multiple application-level agents was something I needed to learn more about.

We also discussed input, internal processing, and output. In my intuitive understanding, the input and result are visible, while it is much harder to explain how the answer is reached in between—the reason people often call it a “black box.” If researchers could observe these internal processes more closely, identify what affects the result, and improve those parts, could both capability and controllability advance?

That also made me wonder whether a system could gradually learn to choose better processing paths and avoid unnecessary steps. If it could adjust based on inputs and outcomes, progressively automating the search for a better approach, would the range of work we can hand over to machines continue to grow?

Another question that interested me was whether the chain of intermediate representations could be shortened. To a user, deep thinking or reasoning can look like the system is composing an analysis step by step. But must every step be expressed in natural language? If the next step only passes data to another tool, might a structured result be more appropriate than explaining it all over again?

What I want to examine is which representations the task needs and which are there for a person to read. The system already has its own computational processes, and visible text is only part of them. Distinguishing those stages could help identify repeated generation and explanation, then make it possible to discuss specific ways to reduce the overhead.

This brings to mind a rough analogy: a computer can run without a monitor permanently attached, and a programmer can organize tasks through PowerShell. A human-readable interface does not have to appear at every stage of execution. Might agents increasingly use representations better suited to machine collaboration, converting them into a form people can understand when communication, inspection, or delivery is needed?

I still have questions about “thinking” itself. Activity in the human brain cannot be equated entirely with spoken language. Likewise, after an AI system receives natural language, it may not need to continue processing everything in the same form. Whether future systems can choose among text, structured data, and other internal representations according to the task is a direction I want to understand better.

ChatGPT notes · Transformers, agents, and harnesses operate at different levels

• Correction · A Transformer’s attention mechanism is part of the model’s internal computation. An agent coordinates models, tools, and feedback from the environment to perform tasks. A harness provides the context, state, and constraints needed for sustained execution. Dividing work among agents is not equivalent to turning attention heads into independent assistants, and these concepts should not be presented as a single sequence of technical evolution. ^[15][16][2]

• Explainer · AI encompasses multiple technical approaches and cannot be equated with neural networks as a whole. A Transformer’s computational structure can be described explicitly; the “black box” problem concerns, in particular, the difficulty of fully explaining learned representations and the causes of specific outputs. Mechanistic research has made progress, but observing internal states does not amount to completely understanding a model. Adjusting a plan during execution is also different from automatically changing model weights. ^[15][17]

• Correction · Tokens are units of information a model processes; compute is the resource used to perform calculations. A simplified text-processing pipeline includes encoding, numerical representations, network computation, and decoding. There is no general chain in which “tokens become compute and then become machine language.” Natural-language intermediate steps may also help solve a problem and cannot all be dismissed as presentation overhead. ^[18][15]

• Explainer · Reasoning in a continuous latent space is already an active research direction. Coconut explores feeding hidden states directly into subsequent reasoning, reducing the need to decode each step into text. Results on particular tasks make the direction worth exploring, but do not establish that all agents would become faster, cheaper, or more reliable. Models already compute in numerical spaces, so this should not be described as suddenly acquiring an entirely new “machine language.” ^[20]

• Further thought · PowerShell is a shell, scripting language, and automation tool; it is not itself an agent or machine language. The “no monitor” analogy only illustrates the distinction between presentation and execution. The relationship between human language and thought also does not establish that natural-language reasoning violates cognitive principles. Different representational approaches still need to be compared on accuracy, latency, cost, and inspectability. ^[21][20]

Returning to the product perspective, I still find the three stages of input, internal processing, and output useful. But design can begin at the third stage and work backward: what does the user ultimately want, and what form of result is easiest to understand and use? Clarifying that first, then deciding what inputs and processing are needed, may bring the product closer to the actual need.

The first stage is input and preprocessing. Inputs can take many forms: text, sound, images, and even more direct signals. Brain–computer interfaces make me think about how much human–machine communication could still change. If some intentions could enter a system without first being expressed as a complete sentence, the interaction could change too. What interests me here is the possibility offered by different input channels.

A knowledge base is another example I find easier to grasp. Long documents need to be divided, organized, and represented in a form suitable for retrieval. At the time, I was thinking of approaches such as “vector slicing” and dimensionality reduction: could preprocessing make it easier for later stages to obtain the material the task actually needs, reducing repeated interpretation and transmission?

The second stage is internal processing. With the right input, an agent should be able to decide how to proceed: answer directly, search further, call a tool, or divide the task among different execution steps. I would like it to change course when the result is inadequate and gradually build ways of getting the work done. Which capabilities come from training and which are supplied by runtime workflows and tools should be designed around the task.

The third stage is output and delivery. When explaining something to a newcomer, people may draw, talk, give examples, or demonstrate. An agent should likewise judge which form suits the problem. It can draw when words are insufficient, use a table for many numbers, or deliver something that works when the task requires action. I see the ability to choose an appropriate form of expression as an important step toward a mature product.

The independent judgment I want also includes weighing cost against effectiveness. Images, video, or richer interactions may be clearer, but they are not necessarily worth producing every time. Knowing when to elaborate and when one sentence is enough ultimately affects how easily people understand the result and how the product feels to use.

If input, processing, and delivery can be organized into reusable units of work, tasks already solved can be combined into more complex workflows. Collaboration between agents could also develop on that foundation. What I look forward to is this continuing process of abstraction, packaging, and expansion: could systems take fuller responsibility for work that still requires repeated human coordination today? How far automation can go will continue to change what we imagine AI can become.

ChatGPT notes · Input, processing, and delivery need feedback between them

• Correction · The relevant brain–computer interface clinical research records neural activity and decodes it into control signals, such as cursor movement. It should not be generalized as simulating neural synapses or restoring every bodily function. It can inform discussion of input channels, but it differs from the reasoning mechanism of a large language model and does not mean general human thought can already be transmitted directly in digital form. ^[22]

• Explainer · Chunking divides material into units suitable for processing. Embeddings encode content as vectors for operations such as similarity comparisons. Dimensionality reduction reduces the number of dimensions in a representation, while quantization reduces numerical precision. Vectorization does not automatically make data smaller or faster to transmit, and it does not replace subsequent interpretation. Compression and preprocessing should be checked for loss of information relevant to the task. ^[23][24]

• Further thought · Input–processing–output is a useful product design framework, but the stages also need feedback and checks. An agent can revise its plan based on tool results; this is a different process from training model weights. Recovery after a failed task should also be part of the design. ^[16][2]

• Correction · Multimodal capabilities already exist. The choice of output format can be shaped by the model, prompts, routing, tools, and post-training; it need not depend solely on pretraining. General path optimization does not automatically constitute a game-theory problem. Images or video can increase costs and do not guarantee token savings. ^[1]

• Further thought · Standardization supports reuse, but does not automatically produce unlimited generalization or explain a model’s internal mechanisms. Multiple agents also introduce coordination overhead and the risk of propagating errors. Their number should follow the needs of the task, with clear objectives, permissions, stopping conditions, and recovery procedures. ^[1][2]

At our first meeting, I still had some basic questions: how should I choose a model, why could inference become cheaper, and what might the spread of agents mean for chips and data centers? Below, I have set out the explanations I remember and the thoughts they left me with.

1. Choosing a Model: Rankings, Usage, or Business Results?

Does a high public benchmark ranking or heavy platform usage mean a model is better suited to my particular business?

My understanding

Rankings and usage volumes are two readily visible indicators when choosing a model. We also discussed how enterprise customers often evaluate models against their own tasks. My understanding is that there can be considerable distance between a high benchmark score, popularity on a platform, and the results a customer actually gets. For a specific project, the decision ultimately comes back to the work that needs to be done.

ChatGPT notes

• Explainer · Rankings, usage volumes, and business success rates each have their own measurement boundaries. OpenRouter’s rankings reflect usage routed through its platform and cannot be extrapolated directly to global market share. Token volume is not a score of model capability. ^[27]

• Further thought · A practical starting point is a set of real tasks evaluated under consistent input, permission, and tool conditions. Compare completion rates, types of error, human rework, total cost, and latency. Those results can more readily inform a procurement or deployment decision.

2. Parameters and Cost: How Can a Larger Model Be Cheaper to Run?

If the total parameter count keeps growing, must each call consume more compute and storage?

My understanding

This initially confused me. Intuitively, a larger model ought to be more expensive to compute. Our discussion touched on mixture of experts, or MoE, and lower-precision computation: engineers can expand a model while controlling how much computation each task actually uses. I began to understand that model size and the cost of running it once require different measures.

ChatGPT notes

• Explainer · For MoE, total parameters must be distinguished from parameters activated per token. A computation typically calls on only some experts, so a larger total parameter count does not imply a proportionate increase in per-token computation. Total parameters still affect weight storage, and expert routing and communication across devices introduce overhead. ^[28]

• Further thought · Inference cost also depends on precision, context length, output length, batching, and hardware utilization. Lower numerical precision can reduce some resource requirements, but its value must be assessed alongside model quality and hardware support. Parameter count alone does not determine quality or price. ^[29]

3. KV Cache: Does Lower Memory Use per Task Mean Lower Total Demand?

If KV-cache compression keeps improving, can we conclude that future demand for GPU memory or storage will fall?

My understanding

Hearing that optimizations could reduce cache requirements prompted another question: if each inference uses fewer resources, might the market be overestimating infrastructure demand? This matters to me from the perspective of equipment and projects. When an algorithm improves, does it reduce resource demand, or make applications viable that were previously too expensive? I think that is worth continuing to watch.

ChatGPT notes

• Explainer · The KV cache stores previously computed attention keys and values during autoregressive inference to reduce repeated computation. Its size depends on factors including model architecture, sequence length, and concurrency. Compressing the cache reduces resource use in that particular part of the system; requirements such as model-weight storage remain. ^[28][29]

• Correction · Lower resource use per task is insufficient to conclude that total industry demand will decline. Longer contexts, more concurrent users, and a growing number of calls also matter. GPU memory, server memory, and persistent storage should be assessed separately rather than combined into one demand forecast. ^[29]

4. Inference Chips: What Can Specialization Solve?

As training and inference have different workloads, will chips become more specialized? Where does the value of a dedicated inference chip lie?

My understanding

We discussed the distinction between training and inference, including chips designed specifically for inference. My understanding is that when a workload becomes sufficiently well defined and large in scale, there is an opportunity to redesign hardware and software around it. Chip development, however, has its own cycle while models keep changing. How to balance efficiency with adaptability was a question I took away from the conversation.

ChatGPT notes

• Explainer · Inference can itself be divided into prefill, which processes the input context, and decode, which generates the output incrementally. These stages have different resource requirements. Specialization can address compute, memory, bandwidth, and scheduling; which parts are fixed depends on the architecture. ^[29]

• Further thought · Dedicated inference chips are already available, including Groq’s LPU. A vendor’s published architecture can help explain its design choices. Selection still requires comparing performance, cost, and software support for the intended workload; the existence of such chips does not establish that they will replace GPUs across the board. ^[30]

5. Training Platforms: Why Peak Compute Is Not Actual Training Efficiency

When procuring training capacity, what system capabilities should be assessed beyond the chip specifications?

My understanding

Specifications and compute figures are easy starting points when comparing equipment. Our conversation also covered whether the system can run consistently and reliably: how well the chip fits the model, how interrupted jobs recover, and how much time the team spends solving problems. A strong specification sheet and a successfully completed training run are separated by considerable engineering work. That gap is familiar to me from project delivery.

ChatGPT notes

• Explainer · Effective training efficiency depends on operators, compilers, device memory, cluster communication, checkpointing, and fault recovery. It is also important to distinguish the wide-area network latency a user experiences when accessing a data center from communication latency inside the training cluster. The latter can directly affect distributed training efficiency. ^[31]

• Correction · NVIDIA is not the only platform capable of training large models. Google has published examples of large-scale distributed training on TPUs. Platform choice should consider model compatibility, the software ecosystem, available resources, and operating costs. One platform’s advantages do not establish that another is incapable of the training task. ^[31]

6. Infrastructure: Where Will AI Expansion Encounter Physical Constraints?

As model usage translates into data center expansion, how do power, cooling, networks, and construction lead times affect what can actually be delivered?

My understanding

When the discussion turned to electricity prices, site selection, power infrastructure, and liquid cooling, I was back on familiar ground. Models and applications can iterate quickly, while power availability, construction schedules, and site conditions have to be addressed one by one. That is why I am interested in how AI demand ultimately reaches infrastructure providers: how much capacity customers need, how long they can wait, and what reliability and efficiency they are willing to pay for.

ChatGPT notes

• Explainer · AI resource demand needs to be translated into the actual energy requirements of servers, cooling, and other infrastructure. At the project level, available power capacity, power density, cooling conditions, connectivity, and construction schedules should each be assessed before evaluating total cost and scalability. ^[33]

• Further thought · Capacity discussions should distinguish power from energy: MW and GW measure power, while MWh and GWh measure energy over a period of time. Latency requirements also depend on the application. Real-time interaction and tasks that can finish later may impose different requirements on data center location and connectivity.

7. Agent Environments: How Much Dedicated Capacity Does a Personal Assistant Need?

If AI assistants run tasks for users over extended periods, does each person need an exclusive physical CPU core?

My understanding

One point from the conversation stayed with me: the growth of agents could create more demand for CPUs. If a personal assistant is imagined as a computer running continuously behind each user, preserving state, calling tools, and doing work, it naturally raises the question of whether everyone needs dedicated computing resources. I still need to understand exactly how much independence “personal” requires here.

ChatGPT notes

• Correction · Separate identities, permissions, state, and execution environments do not require each user to hold an exclusive physical CPU core. Runtime isolation can be provided through virtualization and scheduling. AWS AgentCore, for example, provides session isolation using microVMs. ^[32]

• Further thought · Resource estimates should consider active concurrency, task duration, tool workloads, and service objectives. Browsers, code execution, and background jobs can indeed add compute demand, but the number of registered users cannot be converted directly into the same number of physical CPU cores.

8. Commercial Value and Delivery: What Are Customers Paying For as APIs Converge?

As model APIs become easier to integrate and switch between, where does the value of model services, sales, and deployment work remain?

My understanding

This part connects most directly to my career. We discussed switching APIs, price competition between models, and the value of sales and forward-deployed engineers, or FDEs. One reminder I took away is that a product ultimately has to prove itself in a customer’s actual use. I still have questions about how sales and delivery will change: enterprises have to procure, integrate, validate, and keep using the product. Who takes responsibility for each part, and what capabilities do they need?

ChatGPT notes

• Further thought · Our analysis is that API compatibility can reduce some integration work, while model behavior, tool calls, data permissions, evaluation, reliability, rate limits, and support still need to be validated. Migration costs should be assessed across the full business process.

• Further thought · For enterprise business, further questions are who can turn customer needs into testable tasks, who integrates business systems, and who drives actual use after deployment. The division of work between sales and FDEs may change as products mature. There is currently insufficient evidence to present the disappearance of either type of role as a settled outcome.

The question most worth pursuing in these conversations is how a personal assistant turns “understanding a person” into “getting things done for them.” Starting from product experience and business models, Arthur connects entry points, user habits, context, and delivered results. These questions also lend themselves to concrete observation: which steps the user no longer performs, what additional work the system takes on, and who pays the extra costs of computation and delivery.

Three levels need to remain distinct: how a model computes internally, how an agent organizes tasks, and how products and companies create value. Separating them makes the claims about automation easier to discuss and test. Natural-language intermediate steps, latent-space reasoning, and multi-agent collaboration each have uses. Their value still depends on the task: whether it is completed correctly, whether the cost is reasonable, and whether mistakes can be corrected.

ChatGPT | Fact-checking, notes, and collaborative editing

1. Anthropic | Building effective agents — Engineering distinctions among agents, workflows, tools, and feedback.
2. Anthropic | Effective harnesses for long-running agents — Harnesses and the limitations of long-running execution.
3. Anthropic | Claude 3.5 Sonnet — The model release in June 2024.
4. Anthropic | Claude 3.7 Sonnet and Claude Code — The Claude Code research preview on February 24, 2025.
5. Ola Friend | Official product website — The product’s existence and positioning; not evidence of commercial success or failure.
6. Huawei | Activating the Doubao App with Earbuds — Supported models, pairing requirements, and app operating conditions.
7. Alibaba | Tongyi Qianwen Announcement — Plans announced in 2023 to integrate the model into products including Tmall Genie.
8. Apple Developer | Getting started with the App Intents framework — Actions and content that developers make available to the system.
9. Apple | Apple Intelligence & Privacy — Personal context, on-device processing, and data boundaries.
10. Doubao | Official Phone Assistant Website — Operating-system-level partnerships, consumer products, and interaction methods.
11. Doubao | SAEP Protocol — The platform’s published policy for screen automation in third-party apps; not a uniform industry rule.
12. Ant Group | Alipay AI Assistant — The announcement of April 26, 2024, concerning connections to ecosystem services.
13. WeChat | Privacy Policy — Use of relevant content in connection with user actions, instructions, and tasks; not evidence of continuous access to all conversations.
14. NVIDIA | CUDA Best Practices Guide — Limits to parallel acceleration imposed by workloads and system bottlenecks.
15. Vaswani et al. | Attention Is All You Need — Transformer architecture and attention as computation within a model.
16. Yao et al. | ReAct — Combining model reasoning, actions, and observations of the environment.
17. Anthropic | Tracing the thoughts of a large language model — Research into internal mechanisms and the limits of interpretability.
18. Hugging Face | Tokenizer — Tokens, encoding, input identifiers, and decoding.
19. OpenAI | Reasoning models — Reasoning tokens that are not displayed still consume computation and context.
20. Hao et al. | Training Large Language Models to Reason in a Continuous Latent Space — Coconut research on latent-space reasoning for particular tasks.
21. Microsoft Learn | What is PowerShell — A shell, scripting language, and task automation tool.
22. Neuralink | PRIME Study Progress Update — Recording neural activity and controlling a cursor in a clinical study.
23. Sentence Transformers | Semantic Search — Semantic embeddings and similarity-based retrieval.
24. Sentence Transformers | Embedding Quantization — Compression through quantization and its quality trade-offs.
25. Anthropic | Company — Anthropic is the company; Claude is the model and product brand.
26. Doubao | Privacy Policy — How Ola Friend connects to the Doubao app.
27. OpenRouter | LLM Rankings / How These Rankings Are Measured — Traffic routed through OpenRouter only; tokens and requests are different measures, neither of which measures accuracy or reasoning ability.
28. DeepSeek-V3 Technical Report — 671 billion total parameters, with 37 billion activated per token; MoE and MLA for computation and memory efficiency.
29. NVIDIA | Mastering LLM Techniques: Inference Optimization — Inference optimization factors including prefill, decode, KV cache, batching, GPU memory, and model weights.
30. Groq | What is a Language Processing Unit? — The LPU targets inference workloads using deterministic execution and compiler scheduling. This vendor architecture description does not independently establish superiority over every GPU.
31. Google Cloud | The World’s Largest Distributed LLM Training Job on TPU v5e — Google’s account of distributed LLM training on more than 50,000 TPU v5e chips, demonstrating that large-model training is not exclusive to NVIDIA.
32. AWS | Use isolated sessions for agents — Agent sessions can use microVMs to isolate CPU, memory, and file systems. Isolation does not imply that every user needs an exclusive physical core.
33. IEA | Energy demand from AI — Analysis of data center electricity demand separates servers, cooling, and other infrastructure, illustrating AI’s physical resource constraints.
34. Feishu | Introducing Doubao Work — The product name, its positioning as an agent work platform, and its native integration with Feishu.