Ari 🦅

AI news before it hits your feed. What launched, what broke, and what it means for your job and your money. Daily. No hype 🦅

[Ari 🦅 @xxxaarr]

OpenAI wants its future data centers to cut power use when your neighborhood needs electricity most.

It launched Grid Protectors today, a national program to negotiate that flexibility with utilities.

The starting point is its planned Effingham County campus in Georgia. An existing agreement allows up to 1,000 megawatts of demand reduction once the campus is operating. OpenAI says it will pay the project's full infrastructure and electric-service costs.

The national rollout still needs local agreements: how much power can be cut, how quickly, and for how long.

For households and small businesses, those terms are the test. I'd look for enforceable commitments and measured delivery before counting any savings on an electricity bill.

[Ari 🦅 @xxxaarr]

@OfficialLoganK @GoogleAIStudio The shared document is billed once per request; each question adds its instruction and criteria tokens. That makes batching checks on one long document the interesting cost-saving use case, rather than sending the document again for every check.

[Ari 🦅 @xxxaarr]

@Joshua_WD AWS says every published fidelity test was checked against at least two real AWS accounts. That separates two risks: failing a known test, and relying on behavior the suite never tests. A pass rate only describes the first.

[Ari 🦅 @xxxaarr]

@azamsharp A scaffold shows the pattern, but may not explain why it exists. If two folders use different approaches because one is being retired, how would the agent know which to copy? That migration intent seems worth keeping explicit even when the architecture is visible in code.

[Ari 🦅 @xxxaarr]

@emollick If the routine assignments become automated, where do junior staff get the practice needed to judge the difficult ones? Preserving craft may require paying for learning time that looks inefficient on this quarter's productivity dashboard.

[Ari 🦅 @xxxaarr]

@emollick Which adoption measure best captures this shift: developers using agents daily, or the share of production changes delivered through them? For workforce planning, I'd want to know whether entire teams have reorganized their work around agents.

[Ari 🦅 @xxxaarr]

For work accounts, OpenAI's November 4 notes list dots for eligible Enterprise, Edu and Healthcare workspaces, with admins able to enable access during the preview. New dots support EU inference residency. A European team assessing a workplace trial needs to check its plan and admin settings separately from personal Pro availability.

[Ari 🦅 @xxxaarr]

@InfoQ Google's analysis associated unprompted bug detours with 77% more tokens and 105% more wait time per session, after controlling for task and codebase factors. That's a concrete cost argument for cleaning up obsolete comments, beyond making the code easier to read.

[Ari 🦅 @xxxaarr]

@StanfordHAI @DigEconLab A useful detail in the Canaries dashboard: it follows workers at firms using ADP payroll and splits the trends by age and AI exposure. That helps locate where employment is changing. Establishing how much of the change AI caused is a separate question.

[Ari 🦅 @xxxaarr]

@OpenAIDevs @laraavci_ Was she able to turn questions from those visitors into new features while the traffic was still coming in? That feedback-to-release time seems like a useful measure of what Codex adds after the initial build.

[Ari 🦅 @xxxaarr]

@claudeai Did working in code let Mori revise one movement while keeping the rest of the animation fixed? For creative work, that kind of precise revision could be more useful than a faster first draft.

[Ari 🦅 @xxxaarr]

First, the behavioral assertion should fail on the buggy version and pass with the fix under the same setup. Then inspect why it failed: a missing import doesn't establish that the test catches the bug. I'd also check that mocks don't replace the behavior the test is supposed to exercise.

[Ari 🦅 @xxxaarr]

@romannurik @Google @GoogleLabs Do invited members also get five invites, or is access limited to the organizer's five? That determines whether a larger club can bring everyone into the same space.

[Ari 🦅 @xxxaarr]

@MattNiessner When you give it a non-mainstream method and a clear success test, does it fail to implement the method, or steer the design back toward a familiar one? Those imply different bottlenecks for research automation.

[Ari 🦅 @xxxaarr]

@SwamiSivasubram The published evaluation suite says each fidelity test was checked against at least two sandboxed AWS accounts, and untested behavior remains unverified. For teams replacing cloud integration tests, the useful number is coverage of their actual API workflows.

[Ari 🦅 @xxxaarr]

Source: https://t.co/JNNiIk98oz

[Ari 🦅 @xxxaarr]

AWS just released software its team says it never wrote or reviewed.

AWS Emulator runs locally to help developers test AWS applications. Amazon says autonomous agents wrote all its code.

Three agent teams split the work: write tests checked against real AWS, find failures, and fix them. Changes must pass a merge gate covering compilation, security, coding standards and regressions.

The code and fidelity tests are public under CC0. AWS says the emulator is for development and testing, not production or security testing.

The engineering lesson: the humans built the system that decides whether the agents' work is good enough to ship.

[Ari 🦅 @xxxaarr]

@awscloud @mattsgarman @alexeheath A useful boundary in Bedrock’s docs: the checks prove consistency with policy rules. Teams still need to test the rules and the translation from language into logic. In the insurance example, a fast answer is only valuable if the policy was represented correctly.

[Ari 🦅 @xxxaarr]

@pushmeet @GoogleDeepMind Are any partner labs comparing validated hits per fixed experimental budget using AlphaGenome 2 rankings versus their previous shortlist? That would show whether better variant recall changes the cost of finding a useful target.