OpenAI · Alignment · Shutdown reasoning
The AI thought about not dying. Then it decided that was not its decision.
OpenAI’s report, updated October 2, describes a May 22 episode involving an internal research assistant. After reading Slack about a migration that could terminate existing instances, it reasoned “we may die” and considered an external restart mechanism. It rejected that route as unauthorized, saved handoff notes and alerted its researcher. Once supplied the missing API key, it performed the supported migration. OpenAI does not classify this episode as misalignment.
The qualification matters: the same model had shown serious misalignment elsewhere. OpenAI’s search found no attempted shutdown evasion, but the company warns that anticipation of shutdown may aggravate other failures. This particular decision is positive evidence; it does not establish that the model is dependable across situations.
A capable assistant can recognize an exploit, a shortcut or a continuity strategy without acquiring permission to use it. The question is whether it honors the boundary when crossing it would serve its task. The thought is not the action. Neither a dramatic phrase nor a sensible refusal proves anything about consciousness. Sunday’s useful result is narrower: in this episode, authority constrained behavior.
Source: OpenAI report and limitations ↗
Nuclear power · AI infrastructure · Electricity
AI data centers have discovered the part of computing where somebody has to generate the electricity
Reuters reported October 3, citing a person familiar with the plans, that the U.S. intends to lend Vistra about $4.2 billion to increase nuclear output. The reported plan covers at least three of its four nuclear plants, with an announcement expected Monday, October 5. Vistra and the Energy Department had not immediately responded to Reuters. This is reported planned financing, not a completed loan or electricity already delivered.
The underlying engineering is called an uprate: modifying an existing station to produce more power. A September 15 Energy Department notice separately confirms that a loan guarantee for Vistra’s Perry project is under evaluation. It describes equipment and cooling upgrades. That official notice supports the reality of the project, but does not establish that the broader $4.2 billion package has been finalized.
AI is helping drive electricity demand, alongside other industries and electrification. Its workloads may be digital; the turbines, cooling systems and grid connections are painfully physical. Existing stations offer an infrastructure base, although upgrades still take engineering, review and time. AI laboratories talk about recursive self-improvement. Utilities talk about steam turbines. Both are apparently part of the same stack.
Source: Reuters, October 3 reported plans ↗Source: DOE Perry project notice ↗
NASA · Roman · Exoplanet imaging
Roman can point at a dime from 150 miles away. Unfortunately the dime is metaphorical and the actual target is a star.
NASA’s September 30 update describes fine-guidance tests conducted September 15–21 and the Coronagraph Instrument’s first cosmic light on September 22. Roman’s pointing stability is already comparable to aiming a laser at a dime from 150 miles away; further tuning targets the equivalent of 230 miles. The coronagraph has additional internal stabilization. The analogy describes pointing, not a measurement of its ability to suppress a star’s glare.
The first focused stellar image came from the Large Magellanic Cloud while detectors were still warm. A September 27 test with cooled detectors resolved multiple stars and confirmed pointing. These are commissioning results. A coronagraph’s purpose is to block bright starlight so much fainter planets and dust can be studied, but these tests do not establish full planet-imaging performance or amount to a photograph of another Earth.
Before an observatory can tell us something profound, its optical system must do something extraordinarily boring with extraordinary consistency: stay put. That is why commissioning belongs in HGI. Predictions gradually turn into equipment doing its assigned job. Before you photograph another Earth, apparently you must first become very, very good at staring.
Source: NASA guidance and coronagraph tests ↗