Adeilson Brito

CTO, former CIO & founder of AxonBrief. Writing about enterprise AI strategy, architecture, governance, operating models and agentic systems.

[Adeilson Brito @adeilsonrbrito]

Multi-model debate can improve reasoning—but consensus may come at the cost of independence. The real question is not how many models agree, but how many independent evidence paths survive the debate. https://t.co/IWYt8I4Hd1

[Adeilson Brito @adeilsonrbrito]

Automation is outrunning the human queue. AI can increase the rate at which organizations discover vulnerabilities, propose changes and complete bounded work, but the capacity to authorize, verify, patch, integrate and accept those outputs still scales through people and institutional processes. Technology leaders should therefore stop treating output volume as the primary productivity measure. The governing metrics are queue growth, time to trusted decision, reviewer load, remediation throughput, escaped defects and the fraction of work that reaches independently accepted completion. The strategic design problem is not how to remove people from every loop; it is how to reserve human judgment for the few boundaries where authority and accountability cannot be delegated safely.

[Adeilson Brito @adeilsonrbrito]

AI can improve the quality of our work while weakening our ability to do it without the tool. The real challenge is not just using AI well, but remaining the principal of the work: able to question, verify, intervene, and stand behind the outcome. https://t.co/a4Wqup1JZ0

[Adeilson Brito @adeilsonrbrito]

https://t.co/r8ZIwqhfYd

I think this video is worth taking a look at.

[Adeilson Brito @adeilsonrbrito]

An agent can complete the task and still leave too little evidence to reconstruct what happened. Context pressure may decide what the agent sees next. It should not decide what remains available to inspect. https://t.co/oUQUdOUlNH

[Adeilson Brito @adeilsonrbrito]

This combo — Dario + Sam + Elon in one afternoon proposing:

- delay capability jumps long enough for evals/alignment to catch
- put outsiders inside the training loop with badge-level access
- try to make that the industry norm so no one defects first

The most repeated objection: If China doesn’t pace, a US lab pact just locks winners in.

Whether that pact holds is not decided on X. It is decided by whoever keeps shipping capability without waiting for the pace car — open weights first, then labs that never signed.

Three CEOs can agree to slow the frontier. They cannot make the rest of the world wait.

[Adeilson Brito @adeilsonrbrito]

I would be more cautious than Amodei is his short-term botnet scenario. A real incident demonstrates a failure class, but it does not prove that internet-scale autonomous attacks or recursive self-improvement are imminent.

The geopolitical section is also unresolved: the US capability lead effectively becomes the pacing budget. If the lead narrows, the incentive to slow down also narrows.

There is also an unavoidable incentive problem: Anthropic benefits commercially from being seen as the cautious frontier lab, so its warnings should be taken seriously but not uncritically.

Amodei may be wrong about the exact timeline, but he is probably right about the direction of travel.

[Adeilson Brito @adeilsonrbrito]

@cloneismin Perfect Min. Tokens removed is a local metric. Cost per trusted outcome captures what actually matters: total cost, retries, latency, failures, and whether the agent ultimately delivers a result we can trust.

[Adeilson Brito @adeilsonrbrito]

Fewer terminal characters did not mean cheaper agent work

Evidence:
Quesma published a September 11 benchmark of RTK, a tool that filters terminal output before coding agents read it. The study ran Claude Code with Fable 5.0 and OpenCode with DeepSeek V4 Pro 0813 across 85 and 89 Terminal-Bench 2.1 tasks respectively, five times with and without RTK, for 1,740 attempts after exclusions. On a task-weighted measure, Fable was 1% more expensive with no clear difference from zero and DeepSeek was 17% more expensive; pass rates fell by one and two points. In one failure mode an unsupported rewritten command looped 339 times. Quesma reports that DeepSeek turns were 7% smaller but 18% more numerous. The work is company-authored, uses specific older runtime versions and offers trajectories on request rather than as an immediately downloadable corpus, so generalization should remain cautious.

Interpretation:
Context compression changes the trajectory, not just the invoice. Removing output can cause extra turns, retries or lost information, while cached input may already be cheap. A local proxy such as characters removed or nominal tokens saved therefore cannot establish end-to-end efficiency. This is another instance of Goodhart's law in agent engineering: optimizing a visible internal quantity can worsen cost per successful outcome.

Strategic implication:
Evaluate context interventions at the workflow level. Measure successful outcomes, total billed cost, turns, latency, retries, cache behavior and failure modes against a stable baseline; segment results by model and task rather than relying on one aggregate. Add loop detection and fail-open or bypass behavior when a compressor cannot handle a command. Approve optimizations only when they improve cost per trusted outcome, not when their own counters report large savings.

[Adeilson Brito @adeilsonrbrito]

Astra and Fable debated AI extinction risk with no assigned roles or debate instructions. They challenged each other, revised their claims, and converged on a sharper question: what does the evidence actually support? https://t.co/ueI8mTnNqe

[Adeilson Brito @adeilsonrbrito]

Perhaps the most consequential scenario here isn’t mass unemployment. It’s a much richer economy in which a large part of the population doesn’t feel richer.

If AI breaks the historical link between productivity growth and broadly shared income growth, that may matter more than the headline GDP number.

[Adeilson Brito @adeilsonrbrito]

As AI makes execution abundant, the next bottleneck is judgment. This article explores decision throughput, authority drift, and why agentic systems may need a pacing layer to keep machine speed aligned with organizational control. https://t.co/k0HlLqclAp

[Adeilson Brito @adeilsonrbrito]

@gdb Testing it right now...

[Adeilson Brito @adeilsonrbrito]

@GeminiApp Yes indeed a good model

[Adeilson Brito @adeilsonrbrito]

@EverettTheWest @DTXNaidu Perfect. Recurrence should be event-driven as well as time-based. The control should compare the approved envelope with the live one—tools, data reach, permissions, memory, workers, and model—and reopen assurance as soon as that difference crosses policy.

[Adeilson Brito @adeilsonrbrito]

@DTXNaidu Yes. The missing control is an expiry condition tied to material state change, not elapsed time. Memory promotion, skill installation, model switching, authority expansion, or worker delegation should automatically reopen the assurance case. Approval must bind to system state.

[Adeilson Brito @adeilsonrbrito]

@EverettTheWest That's it. That is the problem I describe as approval decay: the approval remains recorded, but its supporting evidence may no longer describe the operating system. Each material transition must therefore carry its own evidence, authority and rollback path.

[Adeilson Brito @adeilsonrbrito]

An agent can be approved at 9 a.m. and become materially different by 3 p.m.—without a new release. This article explores approval decay and why assurance must follow runtime change. https://t.co/TajrznWJAR

[Adeilson Brito @adeilsonrbrito]

https://t.co/z8nm2VWUyS

[Adeilson Brito @adeilsonrbrito]

A useful but unofficial map of ChatGPT Work's tools.

This is a complete snapshot of the callable interfaces and reusable workflow definitions available to this Work session by Willison.

Link below.