Idea #1
Token generation speed is not primarily a capability lever — it is an ergonomics and attention lever, and the two are easy to confuse.
At current speeds (with 30–45 minute task completion windows), the productive pattern for a solo developer is to kick off 10–15 parallel agents and context-switch between them, which imposes substantial cognitive overhead and fragments attention.
At ultra-fast speeds (~14x faster than the already-fast tier), the math changes: you might run 3–4 agents concurrently, staying genuinely in the loop with each rather than managing a fragmented queue.
More importantly, at real-time speeds combined with voice, you stay in the *flow* of ideation — you can see a prototype generated, steer it, and iterate without ever fully leaving the problem.
The bottleneck also physically relocates: at very high token throughput, the model is no longer the slow part; network round-trips, tool-call overhead, and any other I/O in the agent trajectory become the constraint.
This means the architectural gains from speed require that the rest of the stack (tool APIs, file I/O, sub-agent orchestration) be optimized in parallel, or you only capture a fraction of the theoretical speedup (3–4x in tool-heavy workloads vs. the full 14x in pure generation tasks like prototyping a website or video game).
Idea #2
There are two fundamentally different categories of agentic AI work, and conflating them leads to poor product and system design.
The first is the personal in-flow agent — deeply tailored to one individual, proactive, capable of raising ideas without being asked, operating across technical and non-technical tasks alike, and maintaining a persistent understanding of who you are, what your team is doing, and what your goals are. The defining characteristic is that it must never break the illusion: context switches, sub-agent scaffolding, skill-file management are all implementation details that should be invisible to the user.
The second category is fully autonomous background automation — systems that handle complex, intelligence-requiring processes without a human in the loop, such as scanning production logs and auto-patching performance regressions, or detecting a security vulnerability and closing the exposure window before a human could even triage it.
These two categories require completely different design instincts: the first optimizes for attention-friendliness, continuity of experience, and personalization; the second optimizes for reliability, verification, and minimal-touch human approval gates.
A concrete example is a cybersecurity scanner that finds a vulnerability and automatically patches it, with a human only approving the highest-risk actions — a system where you are not the primary actor but a final safety gate.
Idea #3
The practical, near-term form of recursive self-improvement is not models writing better models — it is frontier models optimizing the infrastructure that serves those models. This is not glamorous but concrete and already delivering results.
The mechanism: the most capable models available are pointed at the inference stack, CUDA kernels, and serving architecture as a software engineering task, and they identify optimizations that a small human team would struggle to find at the same speed.
One of the reported outcomes was a 60% throughput speed improvement over three months and an 80% price reduction using Luna — achieved without changing the model itself, just by re-engineering how it is served.
The key insight is that this is a closed loop: better infrastructure allows more users and use cases, which generates more usage signal, which funds more capability research, which produces more capable models to run the next round of infrastructure optimization. Treating inference engineering as a research-grade problem is the organizational bet that makes this flywheel spin.
Idea #4
Interface naturalness is often the actual adoption inflection point — and voice is the current illustration of this.
When a new, more natural interaction mode becomes available, users move to it faster than capability improvements alone would predict, because humans gravitate to the path of least resistance.
The ChatGPT voice usage data showed this: after launching an improved voice mode (more natural, capable of tool use), the share of users interacting exclusively via voice grew very fast. The practical upshot for practitioners building on top of AI: capability improvements that unlock a more natural interaction pattern (voice, ambient vision, canvas co-editing) tend to expand the addressable audience and usage frequency more than equivalent capability improvements that stay within an already-familiar interaction mode.
The dictation example is for instance dictating morning tasks to ChatGPT, rather than typing them, is not a capability breakthrough (the underlying models area already very powerful) but a friction reduction that changes daily behavior. The implication is that when evaluating new modalities, the friction reduction over the raw capability delta must be kept into consideration.
Source: How to Understand the Next Wave of AI Before Everyone Else | Tibo Interview (Matthew Berman)
Text summarized and optimized using Anthropic’s models and reviewed by a human.