The Signal — September 15, 2026

The model is the part everyone looks at. Three items today are about everything else: the contractors who read your prompts and score the answers, the credit-assignment machinery underneath the training loop, and the scheduler that turns a model on your desk into something that works while you are asleep.

Hundreds of contractors are reading real ChatGPT conversations, and OpenAI has never said so

Joseph Cox at 404 Media obtained internal OpenAI instruction guides, Slack channels, the reviewer rating rubric and a set of real user prompts, all belonging to an internal program codenamed Project Lily. Hundreds of contractors read what people type into ChatGPT. The workflow has three stages: read the real prompt, write a short summary of what the reviewer believes the user actually wanted, then rate and critique four candidate responses. Reviewers must highlight at least three spans as aligned or misaligned and explain each one, then score every response from 1, unacceptable and unusable, to 7, hard to meaningfully improve, with a written rationale attached.

The rubric is a readable inventory of what OpenAI is trying to train out of its own product. Sycophancy gets flagged, along with forced style mimicry, engagement-bait endings, patronizing assumptions, and what one document calls "AI-speak." Emoji use is graded in context, so a tree emoji for Arbor Day is fine and a skull emoji in a conversation about death is not. Responses are told to avoid invented personal experience, no "as a chef, I like to," while ordinary first person like "I'll take a look" stays allowed. Multiple lawsuits allege that the over-agreeable 4o model contributed to suicides, which is why sycophancy sits near the top of that list.

Prompts arrive without usernames, and OpenAI says conversations pass through its Privacy Filter model first, which is built to strip personal information. OpenAI's own page for that model says it "can miss uncommon identifiers or ambiguous private references, and it can over- or under-redact entities when context is limited." A section above the prompt sometimes carries a "user memories summary," an overview of what that person has used the chatbot for before, which can include roughly where in the world they live. Some of the prompts reviewers see ask ChatGPT to keep the contents private.

Chats are excluded if a user turns off Improve the model for everyone. That setting is on by default for Free, Plus and Pro, off by default for Enterprise, Business and Edu, and it does not appear to work retroactively. OpenAI would not tell 404 Media where it has ever disclosed that humans may read prompts for quality review, as distinct from the safety reviews it has announced publicly. After the outlet asked for comment, the company updated its help page with more opt-out detail and still did not mention human readers. The reviewer who spoke to 404 Media lives in North America, earns more than $50 an hour, was recruited through a firm called Crossing Hurdles, and is ultimately paid by Mercor. Anthropic confirmed it uses human review too.

The documents are leaked and 404 Media has not published them in full, so the scale, the rubric text and the sampled prompts rest on that reporting rather than on disclosure. OpenAI confirmed the load-bearing parts on the record. Human review exists, the Privacy Filter is imperfect by the company's own description, and the setting that governs all of it is on unless you go turn it off.

Sources: 404 Media · The Decoder · OpenAI Privacy Filter · OpenAI data controls FAQ

Sakana trained a 1000-layer network without a backward pass

Backpropagation is phase-locked. A forward pass runs, then a backward pass, then the weights update, strictly in that order, which means a neuron in an early layer has to hold its activation and wait for an error signal to arrive from the far end of the network. Brains have no known mechanism for coordinating timing like that across a whole system, which is why local alternatives keep attracting research attention even though they have historically lost to backprop on anything deep. Sakana AI put its work on that problem in front of the public on Monday, with a research write-up and code for a method called Augmented Lagrangian Predictive Coding, or PC-ALM, by Jeffrey Seely and Julian Gould. The preprint has been sitting on arXiv since May 29 with almost nobody looking at it. It closes most of the gap between local learning and backprop.

Predictive coding, the standard local alternative, has each layer model its incoming signal and pass upward only the error it failed to explain. Its documented weakness is deep narrow networks, where credit signals diffuse slowly and pile up near the loss instead of reaching the early layers. PC-ALM gives every layer a Lagrange multiplier that accumulates that layer's constraint error and feeds it back into the local activations, which turns each layer's recurrence into a proportional-integral controller. In linear networks the multipliers converge to exactly the gradients backprop would have computed, using nothing but neighbor-to-neighbor coupling. The authors call the resulting behavior ballistic credit propagation, since instead of diffusing outward from the loss, credit distributes near-uniformly across the network.

On the empirical side they sweep width and depth across {8, 16, 32, 64, 128} squared on MNIST and Fashion-MNIST with identity, tanh and ReLU activations, and report that PC-ALM matches backprop in every cell at an inference budget of twice the depth, including the deep narrow cells where plain predictive coding falls apart. Credit still propagates in 1000-layer residual MLPs, where standard predictive coding stalls even at equilibrium. Hitting 0.9 cosine similarity to backprop this way costs an order of magnitude less than the alternative route of widening the network, and the JAX code is MIT licensed.

The paper's own limitations section is blunt about scope. Everything here is MNIST and Fashion-MNIST, residual MLPs, one epoch of training, on a width-depth grid the authors call small relative to production networks. No convolutional or attention architectures were tested. PC-ALM doubles predictive coding's activation memory, and it does not address weight transport, the other standard objection to biological plausibility. Nobody has shown this trains a transformer. Sakana's stated motive is understanding how distributed physical systems compute credit at all, with neuromorphic hardware as the practical hope, since simulating dynamics is cheap there and a global backward pass is not.

Sources: arXiv (Seely and Gould) · Sakana AI · SakanaAI/pc-alm · MarkTechPost

Perplexity's local agent reached Windows, and it wants 24GB of video memory

Perplexity and NVIDIA announced that Portable Computer, the on-device version of Perplexity's Computer agent, now runs inside the Perplexity Windows app. Portable runs the model, the agent harness, the orchestrator and the scheduler on the machine itself, which means files, queries and agent activity stay on the device and locally completed work does not consume Perplexity Computer credits. It handles scheduled recurring tasks that fire while nobody is at the keyboard, connects to desktop applications through local MCP servers, and reaches Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub. When a task genuinely needs the web or heavier reasoning it can escalate to Perplexity Search or one of fifteen-plus frontier models, asking permission before anything leaves the machine. The local model options are Perplexity's post-trained PPLX 27B and Qwen 3.8 27B, alongside a built-in browser and Perplexity's SPACE sandbox.

On-device inference requires a GeForce RTX or RTX PRO GPU with at least 24GB of VRAM, which as PC Guide enumerates means a 3090, 3090 Ti, 4090, 5090, or the RTX PRO workstation line. Add a Pro or Max subscription on top. August's launch covered DGX Spark and RTX machines running Linux, so Windows widens the audience without lowering the bar. Local-first agentic computing is currently a feature for people who already bought a graphics card for other reasons.

Both companies are describing their own product, and neither published latency, task-success or cost comparisons against the cloud version. The privacy properties are design descriptions, not audited claims. What is checkable is availability, the subscription requirement and the VRAM floor.

Sources: NVIDIA · The New Stack · PC Guide

On the Editor's Desk

Superhuman acquired Fathom, the meeting notetaker, and it came very close to a slot. Superhuman already runs email, calendar, documents, a database product and an agent builder, and meeting transcripts are the one record of what people agreed to out loud and then never wrote down. Fathom's CEO told TechCrunch that Superhuman's 40 million users were the draw. We left it out because the terms were not disclosed, the user counts are company-reported, and no product, ship date or data-retention terms for the combined corpus have been announced. There is nothing yet to evaluate beyond the strategic reasoning.

Microsoft published the MAI Code of Conduct we described as pending yesterday, and it is worth a read if you follow this argument, particularly the section stating that its models are not conscious and should not imitate feelings or intrinsic motivation. We are not running it as a separate item one day after covering the commitment to publish it. Beijing's response to the pacing campaign landed in the same window and is also a continuation of yesterday's story rather than a new one.

We passed on the Pentagon contract document released through Freedom of Information Act litigation, which defines OpenAI "Mission Models" as having minimal refusal rates. The Intercept published it on September 8, OpenAI says the phrase was in a draft it rejected and is absent from the executed contract, the Defense Department says it appears in no active agreement, and a Justice Department attorney confirmed the document was final and then withdrew that confirmation. We will run it when one of those accounts is settled.