Matthew Berman · Published 2026-08-24

How to Understand the Next Wave of AI Before Everyone Else | Tibo Interview

privateReview: unreviewedConfidence: high
inferenceagentscodexefficiencyopenaiHarnesslatencymulti-agentcost-efficiencycloud-agentsorchestration
Open on YouTube ↗

Summary

Overview

  • Speaker: Matthew Berman and Tibo
  • Channel: Matthew Berman
  • Main topic: The future of AI, model efficiency, agent workflows, and the evolution of AI development culture
  • Purpose: To provide deep insights into the evolution of AI models, development culture, efficiency gains, and agentic workflows from an insider perspective. Matthew Berman interviews Tibo, discussing his background on the Google DeepMind team working on early language models like LMChat before ChatGPT, transitioning to OpenAI, the importance of speed (ultrafast models), agent harnesses, and the cultural philosophy of shipping products quickly and leaning into natural human-AI interaction.

Topic Map

Google DeepMind and LMChat Background

  • Explanation: Tibo discusses working on LMChat at DeepMind a year before ChatGPT, why Google was hesitant to release it, and the cultural differences between DeepMind and OpenAI regarding shipping products.
  • Key claims:
    • Google was too nervous to release LMChat and blocked DeepMind from shipping disruptive products.
    • OpenAI has a strong bias to ship and make things available to people quickly.
    • DeepMind was a creative, research-focused place with less emphasis on shipping production products.
  • Examples:
    • Working on language models and scaling them at DeepMind prior to ChatGPT's public release.
  • Terminology:
    • LMChat
    • DeepMind
    • OpenAI
    • codenames
  • Why it matters: Highlights how institutional risk aversion and cultural bias toward shipping shape the competitive AI landscape.

Model Efficiency, Speed, and Ultra-Fast Inference

  • Explanation: Exploration of how token generation speeds (e.g., 750 tokens per second) transform user workflows, reduce cognitive overhead, and enable parallel agent execution.
  • Key claims:
    • Ultra-fast speeds (750+ tokens per second) will become the default within a year or two.
    • High token speeds change the bottleneck from bandwidth to CPU tool calls.
    • Fast models allow running multiple agents in parallel without cognitive fatigue from context switching.
  • Examples:
    • Using voice models and transcription for efficiency instead of typing prompts.
    • Running 10 to 15 coding agents in parallel with ultrafast speeds.
  • Terminology:
    • tk/s
    • ultrafast
    • token speed
    • latency
    • compute capacity
  • Why it matters: Speed fundamentally changes how humans interact with AI, turning tools from slow query engines into fluid collaborators.

AI Culture, Conviction, and Disruption

  • Explanation: Discussion on maintaining a bottom-up, empowering culture at OpenAI where employees can push ideas and disrupt existing products without heavy friction.
  • Key claims:
    • OpenAI has a bottoms-up, empowering culture with very little stop energy for new product ideas.
    • Founders must have conviction, listen to user feedback, and be willing to disrupt their own products.
    • Balancing simplicity and high quality while empowering rapid iteration is critical.
  • Examples:
    • Shipping Codex, ChatGPT iOS app, and new voice features rapidly based on internal and user feedback.
  • Terminology:
    • stop energy
    • bottoms-up culture
    • conviction
    • iterating
  • Why it matters: Internal culture determines whether an organization can outpace competitors by cannibalizing its own successful products.

Agentic Coding, Harnesses, and Workflows

  • Explanation: Examining how coding agents (like Codex) and agent harnesses will evolve beyond simple command-line tools into sophisticated multi-agent ecosystems.
  • Key claims:
    • Current coding harnesses will seem primitive in 2-3 months as models require more than a laptop.
    • Cloud agents and automated loops reduce the need for humans to be in every step of the loop.
    • The ideal interface adapts to individual human preferences rather than forcing a rigid coder or non-coder mode.
  • Examples:
    • Codex reaching 20 million users and experiencing explosive vertical growth.
    • Using automated loops for code generation, testing, and compilation concurrently.
  • Terminology:
    • harness
    • codex
    • cloud agents
    • loops
    • context switching
  • Why it matters: Agentic workflows shift software development from writing individual lines of code to managing orchestrations of AI agents.

Competition, Pricing, and the Future of AGI Interfaces

  • Explanation: Comparing OpenAI and Anthropic, discussing pricing drops (like Luna/Claude pricing updates), and envisioning personalized personal AI assistants.
  • Key claims:
    • OpenAI focuses on uniquely well-done value and maximum acceleration rather than watching competitors.
    • Algorithmic efficiency gains and strategic compute planning drive massive price drops while maintaining margins.
    • The end state is a personalized personal AGI interface deeply tailored to each individual's unique workflow.
  • Examples:
    • Price drops on Luna and similar models by 80% due to efficiency gains.
    • Using voice interaction and ambient integration as natural extensions of human communication.
  • Terminology:
    • luna
    • algorithmic efficiency
    • personal agi
    • multimodal
    • margins
  • Why it matters: Shows how aggressive hardware and algorithmic scaling will democratize access to advanced intelligence globally.

Key Points

Speed changes cognitive workflows

  • Explanation: When inference speeds increase significantly, users can run multiple agents in parallel and stay in a continuous creative flow.
  • Evidence: Ultra-fast token speeds allow 10-15 parallel agents without excessive cognitive overhead.
  • Practical implication: Developers and creators can delegate complex, multi-step tasks simultaneously rather than sequentially.

Culture of shipping over risk aversion

  • Explanation: Organizations that hesitate to ship out of fear of disruption lose ground to those with a bias toward action and deployment.
  • Evidence: Google's hesitation with LMChat vs. OpenAI's rapid deployment strategy.
  • Practical implication: Companies must build cultures that embrace self-disruption and fast iteration cycles.

Recursive self-improvement and infrastructure

  • Explanation: Efficiency gains come from both algorithmic improvements and optimizing the underlying inference stack recursively.
  • Evidence: Substantial price drops and 60%+ speed increases achieved over short multi-month periods.
  • Practical implication: Continuous compute planning and infrastructure optimization yield compounding performance advantages.

Frameworks, Models & Processes

The Two-Category Problem Framework

  • How it works: Categorizing AI problem-solving into personal hyper-tailored agents versus fully automated systemic workflows.
  • Components:
    • Personalized Proactive Agents
    • Full-On Automation Systems
  • When to use: When designing AI products to serve either individual daily workflows or complex enterprise-grade automated pipelines.

Examples & Case Studies

Google developed LMChat prior to ChatGPT but restricted its release due to internal fears of disrupting search.

  • Illustrates: The dangers of institutional risk aversion in fast-moving technology fields.
  • Lesson: Ship early and embrace disruption before others do it for you.

Codex usage grew rapidly, transitioning from a primitive coding harness to a massive developer tool with millions of users.

  • Illustrates: The exponential adoption curve of AI-assisted software development tools.
  • Lesson: Next-generation developer tools must transcend simple local environments to leverage cloud scale.

Actionable Takeaways

  • Immediate:
    • Leverage ultrafast inference models to run parallel tasks and maintain workflow momentum.
    • Embrace voice interfaces and multimodal prompts to reduce typing friction.
  • Strategic:
    • Foster a corporate culture with low stop energy and a strong bias toward shipping.
    • Plan compute capacity aggressively to capture efficiency gains and pass cost savings to users.
  • Questions to investigate:
    • How will solo developer workflows evolve as agent harnesses become fully autonomous?
    • What are the long-term societal impacts of ubiquitous personal AGI assistants?

Claims Worth Verifying

  • Google had LMChat ready a year before ChatGPT and blocked its public release. (historical fact)
  • Token generation speeds have increased by roughly 60% over three-month periods through architectural optimization. (technical metric)

Notable Quotes

"I can press the button whenever I want, whenever it feels right." (at 0:00) "Google was too nervous to release it and DeepMind was blocked from shipping products that could disrupt Google." (at 0:57) "We are a very, very forward looking, and the the the future of AI and what it will all look like and how humanity benefits... doesn't really wait." (at 9:05)

Compressed Summary

  • Early work on LMChat at DeepMind and why Google delayed deployment.
  • The transformative impact of ultrafast inference speeds (750+ tk/s) on workflows.
  • OpenAI's culture of bottom-up empowerment and bias toward shipping.
  • The evolution of agentic coding harnesses and multi-agent concurrency.
  • Balancing algorithmic efficiency gains, price drops, and compute planning.
  • Keywords: inference, agents, codex, efficiency, openai
  • Core insight: Rapidly shipping highly efficient, multimodal AI models and fostering an empowering culture of self-disruption are the keys to leading the next wave of technological evolution.

Core insights

7
Mechanismhigh noveltymoderate evidence

At ultra-fast token rates (~750 tok/s), the binding constraint in an agent loop migrates from model inference to CPU-bound tool execution, so the harness/tool-runner becomes the critical path that must be scheduled and parallelized rather than the model call itself.

Why it matters

It relocates optimization effort: teams that keep tuning model latency/throughput will plateau, while tool fan-out, sandbox provisioning, and host CPU contention become the dominant latency contributors in end-to-end agent runs.

Generalization

Whenever one stage of a pipeline is accelerated by orders of magnitude, the adjacent stage inherits the bottleneck; architecture should be re-profiled, not just re-tuned, after a step-change in component speed.

Ultra-fast speeds (750+ tokens per second) will become the default within a year or two.
Open source video
High token speeds change the bottleneck from bandwidth to CPU tool calls.
Open source video
Practicemedium noveltymoderate evidence

Scaling to 10-15 concurrent coding agents is framed as a human-supervision problem, not a compute problem: the limit is the operator's working-memory cost of context switching between agent streams.

Why it matters

It implies the highest-leverage investment in multi-agent systems is supervision UX and state visibility (interrupts, diff review, per-agent status) rather than raw parallelism, and that observability tooling is the real scaling lever.

Generalization

Human attention becomes the scarce resource in agentic systems; architecture should expose succinct, interruptible state per concurrent worker rather than raw log streams.

Fast models allow running multiple agents in parallel without cognitive fatigue from context switching.
Open source video
Running 10 to 15 coding agents in parallel with ultrafast speeds.
Open source video
Predictionhigh noveltyweak evidence

Coding harnesses are treated as a rapidly depreciating abstraction: current harnesses are expected to look primitive within 2-3 months because models will need more compute than a laptop, pushing execution into cloud agents.

Why it matters

It argues against deep investment in long-lived local-first harness abstractions and for harness designs where the execution substrate (local vs cloud) and model are swappable behind a stable interface, with remote-first session state and credentials.

Generalization

Tools that bundle model + runtime + local execution have short half-lives when model requirements grow faster than developer machines; keep the orchestration contract stable and the execution backend replaceable.

Current coding harnesses will seem primitive in 2-3 months as models require more than a laptop.
Open source video
Cloud agents and automated loops reduce the need for humans to be in every step of the loop.
Open source video
Architecturemedium noveltymoderate evidence

Agent loops are moving from sequential human-gated steps to concurrent automated loops (generation, testing, compilation running together), which changes the unit of work from a single code edit to a self-scheduling pipeline.

Why it matters

Concurrent generate/test/compile loops require idempotent builds, isolated environments per loop, and automated rollback or gating, otherwise concurrent loops contend on shared state and produce nondeterministic failures.

Generalization

Increasing autonomy inside an agent converts orchestration problems into distributed-systems problems: isolation, concurrency control, and determinism become the design constraints.

Using automated loops for code generation, testing, and compilation concurrently.
Open source video
Empirical Resultmedium noveltymoderate evidence

Large price reductions (cited at ~80%) are attributed to algorithmic efficiency gains plus strategic compute planning sustained while preserving margins, which is framed as the mechanism that makes always-on and parallel agent workloads economically viable.

Why it matters

Cost per token determines whether continuous background loops, multi-agent fan-out, and ambient assistants are viable products; cost curves, not capability curves, gate these architectures.

Generalization

Provider economics are driven by algorithmic efficiency rather than only hardware scale, so cost-sensitive agent architectures should be designed to ride expected unit-cost declines and to re-benchmark as prices fall.

Price drops on Luna and similar models by 80% due to efficiency gains.
Open source video
Algorithmic efficiency gains and strategic compute planning drive massive price drops while maintaining margins.
Open source video
Mental Modelmedium noveltyweak evidence

The proposed end state of agent interfaces is a per-user adaptive interface rather than a fixed coder/non-coder mode split, plus natural modalities such as voice/transcription replacing typed prompts.

Why it matters

It implies interface personalization and modality selection belong inside the agent runtime (learned per user), not as a static product-mode toggle, which changes where user preference state and intent-capture logic live.

Generalization

As capability ceilings rise, differentiation shifts from capability exposure to intent capture: minimizing the friction between the human's intent and the agent's action.

The ideal interface adapts to individual human preferences rather than forcing a rigid coder or non-coder mode.
Open source video
Using voice models and transcription for efficiency instead of typing prompts.
Open source video
Failure Modelow noveltymoderate evidence

Risk aversion is presented as a concrete competitive failure mode: an organization holding back a shippable model out of fear of disrupting its own core product lost the timing advantage to a competitor with a bias to ship.

Why it matters

It frames internal 'stop energy' as a measurable latency tax on deployment, which matters for teams deciding whether to release a product that cannibalizes their existing surface.

Generalization

Organizations optimizing for protection of existing revenue streams systematically under-ship in platform transitions; self-cannibalization is a strategy, not a bug.

Google was too nervous to release LMChat and blocked DeepMind from shipping disruptive products.
Open source video
Founders must have conviction, listen to user feedback, and be willing to disrupt their own products.
Open source video

Deep dives

5

Latency decomposition of agent loops under ultrafast inference

Research question

At 750+ tokens per second, what fraction of end-to-end agent-loop latency and cost is attributable to CPU-bound tool execution (dispatch, sandbox setup, host execution) versus model inference?

Why

Determines whether harness engineering or model-serving optimization is the higher-leverage investment; if tool time dominates, faster models deliver diminishing returns unless the tool path is redesigned.

High token speeds change the bottleneck from bandwidth to CPU tool calls.
Open source video
Ultra-fast speeds (750+ tokens per second) will become the default within a year or two.
Open source video
Source video

Human supervision ceiling for concurrent agents

Research question

How many concurrently running agents can a single operator supervise before review quality and error detection degrade, and does supervision UI design shift that ceiling?

Why

Determines the practical parallelism limit of multi-agent coding systems and whether investment belongs in supervision UX rather than raw concurrency.

Fast models allow running multiple agents in parallel without cognitive fatigue from context switching.
Open source video
Running 10 to 15 coding agents in parallel with ultrafast speeds.
Open source video
Source video

Harness half-life and execution-substrate swappability

Research question

What is the realistic half-life of a coding-harness abstraction, and what interface keeps the harness stable while the model and execution substrate change?

Why

Prevents stranding investment in local-first harness designs that become obsolete as model compute requirements outgrow developer machines.

Current coding harnesses will seem primitive in 2-3 months as models require more than a laptop.
Open source video
Cloud agents and automated loops reduce the need for humans to be in every step of the loop.
Open source video
Source video

Concurrency control for automated generate/test/compile loops

Research question

What isolation, idempotency, and gating mechanisms are required for concurrent generate/test/compile loops to avoid nondeterministic failures and shared-state contention?

Why

Converts multi-agent autonomy into a distributed-systems design problem, where isolation and determinism become the limiting constraints.

Using automated loops for code generation, testing, and compilation concurrently.
Open source video
Source video

Decomposing unit-cost declines into algorithmic, hardware, and utilization factors

Research question

How much of the reported ~80% price decline comes from algorithmic efficiency versus hardware or compute-planning effects?

Why

Cost trajectories determine the viability of always-on and fan-out agent architectures; misattributing declines leads to wrong capacity planning.

Price drops on Luna and similar models by 80% due to efficiency gains.
Open source video
Algorithmic efficiency gains and strategic compute planning drive massive price drops while maintaining margins.
Open source video
Source video

Article ideas

4

The Bottleneck Moves: Why Ultrafast Models Make Your Harness the Critical Path

Ultrafast inference does not make agent loops fast; it relocates the critical path to CPU-bound tool execution, so harness engineering must converge with job-scheduler and runtime engineering, and teams that keep tuning model latency will plateau.

Angle

Performance engineering consequence for agent builders

Source video

Ten Agents, One Human: The Real Ceiling on Multi-Agent Coding Is Supervision, Not Compute

The practical limit on parallel agents is operator working-memory bandwidth, so the highest-leverage investment is supervision UX — per-agent status, interruptibility, diff review — rather than raw concurrency; token speed alone does not remove context-switching cost.

Angle

Human-factors counterpoint to parallelism hype

Source video

Your Harness Has a Half-Life: Designing for Model and Substrate Swaps

Because frontier coding models will require more compute than a laptop, local-first harness abstractions get stranded; the durable design is a stable orchestration contract over swappable remote execution substrates and model backends.

Angle

Architecture pattern for durability under model churn

Source video

The Cost Curve Is the Capability Curve: Architecting Agents Against Unit-Cost Declines

Always-on and multi-agent fan-out architectures are gated by cost per token, and provider price declines come from algorithmic efficiency and compute planning rather than hardware alone; agent architectures should therefore be designed to ride expected unit-cost declines and re-benchmark as prices fall.

Angle

Economics-driven architecture argument

Source video

Project ideas

4

LoopLatency Decomposer

new

At ≥750 tok/s, CPU/tool execution accounts for the majority (>50%) of end-to-end agent-loop latency on representative coding tasks, and its share increases monotonically with token speed.

Proof of concept

Instrument an open coding harness with per-stage timers (model streaming, tool dispatch, sandbox provisioning, host execution) and sweep token-throughput tiers by swapping model providers or configuring decoding speed.

Measurement

Share of loop latency in tool execution vs model inference per task, and the regression slope of tool share against token speed; cost per loop decomposed the same way.

Source video

Supervision Ceiling Testbed

beyond-evals

Per-agent review quality degrades beyond a small number of concurrent agents (e.g., >5) independent of token speed, and a diff-centric, interruptible status UI raises the ceiling versus raw log streams.

Proof of concept

Run scripted agent task streams with planted defects under controlled agent counts, comparing a raw-log supervision surface against a per-agent status/diff/interrupt surface.

Measurement

Missed-defect rate, time-to-intervention, and subjective load as functions of agent count and UI design; ceiling N where missed-defect rate exceeds a threshold.

Source video

Harness Migration Cost Benchmark

new

A harness with a fixed orchestration contract and swappable model/execution backends incurs lower migration cost (engineer-hours and behavioral regressions) across consecutive model generations than a model-bundled local-first harness.

Proof of concept

Implement two thin coding harnesses — contract-first versus bundled — and port both across two model backends and two execution substrates (local vs cloud), measuring the change surface.

Measurement

Lines changed, engineer-hours, and count of behavioral regressions per port; time-to-first-successful-run on the new backend.

Source video

Concurrent Loop Isolation Runner

movement-lab

Concurrent generate/test/compile loops without per-loop isolation and idempotent builds produce nondeterministic failures and shared-state contention that scale superlinearly with loop count; isolation plus automated gating eliminates them.

Proof of concept

Build a loop runner that provisions isolated sandboxes per loop, enforces idempotent build artifacts, and adds automated merge/rollback gating; run N concurrent loops on a repo with and without isolation.

Measurement

Nondeterministic failure rate, shared-state conflict rate, time-to-merge, and rollback frequency as a function of loop count and isolation mode.

Source video

Architectural implications

4

Inference speed stops being the dominant latency term in agent loops.

Before

Optimize the model path: quantization, batching, streaming, provider selection.

After

Optimize the tool path: concurrent tool dispatch, sandbox warm pools, host CPU budget, and result serialization.

Consequence

Harness engineering converges with job-scheduler/runtime engineering; profiling must decompose end-to-end loop latency into model vs tool-call time.

Source video

Execution migrates from the developer's laptop to cloud agents.

Before

Harness runs locally with local filesystem, local credentials, and local model inference.

After

Harness is a thin client over a remote execution environment with remote session state and scoped credentials.

Consequence

Session persistence, credential scoping, and reproducible remote environments become first-class requirements; local-only harnesses become dead ends.

Source video

Humans shift from a step inside each loop to a supervisor of many concurrent loops.

Before

Human approves or edits at each step of a single sequential workflow.

After

Human monitors and intervenes across many autonomous loops running concurrently.

Consequence

Supervision surfaces (status, diff review, interrupts, merge/rollback controls) become the primary product surface and the practical limit on parallelism.

Source video

Interface mode is personalized per user rather than fixed by product mode.

Before

Separate coder and non-coder modes with different affordances.

After

A single adaptive interface that infers modality and abstraction level per user.

Consequence

Per-user preference state and intent-capture telemetry must live in the runtime, raising privacy and evaluation requirements for personalization.

Source video

Tradeoffs and failure modes

5

Ultrafast inference shifts the bottleneck rather than removing it

Benefit

Fluid, near-instant interaction and heavily parallel agent execution.

Cost or risk

CPU-bound tool calls become the limiting factor, so gains from faster models are partially absorbed by tool execution overhead unless the harness is redesigned.

High token speeds change the bottleneck from bandwidth to CPU tool calls.
Open source video
Source video

Rapid harness obsolescence

Benefit

Fast iteration on agent capability; each generation of harness unlocks new model behavior.

Cost or risk

Investment in local-first harness abstractions can be stranded within months as models outgrow laptop-class compute, favoring cloud execution.

Current coding harnesses will seem primitive in 2-3 months as models require more than a laptop.
Open source video
Source video

Self-cannibalization vs protecting existing revenue

Benefit

Shipping disruptive products first captures the transition and prevents a competitor from doing so.

Cost or risk

Releasing a model that undercuts the incumbent product risks damaging the existing business, which is exactly the pressure that delays shipping.

Google was too nervous to release LMChat and blocked DeepMind from shipping disruptive products.
Open source video
Source video

Aggressive price cuts enabled by efficiency

Benefit

Democratized access and economically viable always-on/parallel agent workloads.

Cost or risk

Sustaining large price drops requires continuous algorithmic and compute-planning gains; providers without them face margin compression.

Algorithmic efficiency gains and strategic compute planning drive massive price drops while maintaining margins.
Open source video
Source video

Human-in-the-loop removal via automated loops

Benefit

Concurrent generate/test/compile loops remove human latency from each iteration and raise throughput.

Cost or risk

Reducing human checkpoints per loop reduces per-step oversight, shifting the safety burden to automated gating and verification.

Cloud agents and automated loops reduce the need for humans to be in every step of the loop.
Open source video
Source video

Open questions

5

At 750+ tokens per second, what fraction of end-to-end agent-loop latency and cost is attributable to CPU tool calls versus model inference?

Why unresolved

The summary asserts the bottleneck shifts to CPU tool calls but provides no measurement or breakdown.

Research direction

Instrument a representative agent harness and produce a latency/cost decomposition per loop across token-speed tiers, identifying whether dispatch, sandbox setup, or host execution dominates.

Source video

How many concurrent agents can a single operator supervise before review quality and error detection degrade?

Why unresolved

The claim that parallelism runs 'without cognitive fatigue' is an assertion from the speaker, not a measured result.

Research direction

Human-factors study varying agent count with controlled review tasks, measuring missed defects and time-to-intervention as a function of supervision UI design.

Source video

What is the realistic half-life of a coding-harness abstraction, and what design keeps the harness stable while the model and execution substrate change?

Why unresolved

The '2-3 months' horizon is a prediction without supporting evidence in the summary.

Research direction

Track harness churn empirically; prototype a harness with a fixed orchestration contract and swappable model/execution backends to measure migration cost across model generations.

Source video

How much of the reported ~80% price decline comes from algorithmic efficiency versus hardware or compute-planning effects?

Why unresolved

The summary attributes price drops to efficiency gains and compute planning but gives no decomposition.

Research direction

Collect public price/throughput series per model generation and decompose cost per token into algorithmic, hardware, and utilization factors.

Source video

Does a personalized adaptive interface outperform an explicit coder/non-coder mode split on task completion and error rates?

Why unresolved

The adaptive-interface end state is stated as a vision rather than an evaluated design.

Research direction

A/B test adaptive modality and abstraction selection against fixed modes, measuring task success, correction frequency, and user-reported friction.

Source video

Key claims

8
predictionVerification needed

Ultra-fast inference speeds of 750+ tokens per second will become the default within one to two years.

Evidence

Ultra-fast speeds (750+ tokens per second) will become the default within a year or two.

Question

What is the current distribution of served token throughput across major providers, and what trajectory would reach 750+ tok/s as a default within two years?

Source video
causalVerification needed

At high token speeds, the performance bottleneck moves from bandwidth to CPU tool calls.

Evidence

High token speeds change the bottleneck from bandwidth to CPU tool calls.

Question

In a measured agent loop at high token throughput, does CPU/tool-execution time exceed model serving time as the dominant latency term?

Source video
causalVerification needed

Fast models allow running multiple agents in parallel without cognitive fatigue from context switching.

Evidence

Fast models allow running multiple agents in parallel without cognitive fatigue from context switching.

Question

Does increasing inference speed measurably reduce operator context-switching cost, or does supervision load scale with agent count independent of token speed?

Source video
predictionVerification needed

Current coding harnesses will appear primitive within two to three months because models will require more than a laptop.

Evidence

Current coding harnesses will seem primitive in 2-3 months as models require more than a laptop.

Question

What compute footprint do current frontier agentic coding workloads require, and at what rate is it outgrowing developer-side hardware?

Source video
factualVerification needed

Codex reached 20 million users with explosive vertical growth.

Evidence

Codex reaching 20 million users and experiencing explosive vertical growth.

Question

What is the verified user count and growth curve for Codex as of the recording date?

Source video
comparativeVerification needed

Model prices dropped roughly 80% (e.g., Luna and similar models) due to efficiency gains while margins were maintained.

Evidence

Price drops on Luna and similar models by 80% due to efficiency gains.

Question

Which specific models saw ~80% price reductions, over what period, and were provider margins actually preserved?

Source video
opinionVerification needed

Google was too risk-averse to release LMChat and blocked DeepMind from shipping disruptive products, whereas OpenAI has a strong bias to ship.

Evidence

Google was too nervous to release LMChat and blocked DeepMind from shipping disruptive products.

Question

What internal decision process and timeline led to LMChat not being released publicly prior to ChatGPT?

Source video
causalVerification needed

Algorithmic improvements plus recursive optimization of the inference stack produced 60%+ speed increases and substantial price drops over short multi-month periods.

Evidence

Substantial price drops and 60%+ speed increases achieved over short multi-month periods.

Question

Can the 60%+ speed improvements be attributed to algorithmic changes, inference-stack optimization, or hardware, and over what exact intervals?

Source video

Connections

5