In partnership with
In Today's Issue:
⚠️ OpenAI's self-generated jailbreak instructions
🎬 Andrew Yang's warning about the internet
🛠️ Claude brings Cowork into chat
🧬 AI-designed treatments for aging immune cells
✨ And more AI goodness…
⚡ The Signal
AI is moving fast enough to make even its builders ask for more time.
OpenAI has published six cases of model misalignment and says the industry cannot responsibly keep scaling at maximum speed for much longer. In today's video, Andrew Yang relays an unsettling account from an unnamed lab leader who believes OpenAI-linked bot swarms have contaminated the internet. Jensen Huang opposes a coordinated slowdown, while DeepMind's new institute opens a public debate about how to prepare for AGI, AI with broad human-level capabilities. Meanwhile, Claude is making it easier to hand work to an agent, and GLM has helped build infrastructure for serving a model. The useful applications and the reasons for slowing down are arriving together.
All the best,
Kim Isenberg
Editing a slide through a comment in Claude (Anthropic, official product demo).
🛠️ Claude Brings Cowork Into Chat
Anthropic is merging Claude chat and Cowork, with the new experience rolling out to Pro and Max over the coming weeks and other plans following. Cloud tasks can continue after you close your laptop; work involving local files or apps needs Claude Desktop open. Context, skills and connectors carry across, while Docs and Slides add document and presentation editing inside the conversation, in beta on paid plans. Claude asks before taking actions by default.
👉 tl;dr: Hand Claude a report and a slide deck in one conversation, without moving between separate workspaces.
🌍 Mozilla Finds Open AI Models Closing In
Mozilla's September 15 report puts the open-model gap at about 4.4 months, measured by the length of tasks models can complete. Using data through September 1, it finds the best open model three points behind the closed leader on the Artificial Analysis Intelligence Index, a combined benchmark score, at 60% of its API price. The hardware chart shows what it takes to run these models yourself, with the strongest performance requiring substantial equipment.
👉 tl;dr: Open models bring near-frontier performance at lower API prices, with the option to run them yourself on suitable hardware.
The DeepMind Institute's introductory essay, by Shane Legg, James Manyika and Demis Hassabis (Google DeepMind).
🧠 DeepMind Opens an Institute for AGI
Shane Legg, James Manyika and Demis Hassabis have introduced the DeepMind Institute, a platform for research and debate about artificial general intelligence, or AGI. Researchers from Google and DeepMind want to bring in outside voices to examine how broadly capable AI should be built, governed and used. Early essays cover reasoning transparency and economic policy. The invitation extends to government, the arts and humanities: help work out what society needs before AGI arrives.
👉 tl;dr: DeepMind is opening the AGI discussion to people who study society, as well as those building the models.
🎬 Watch This
Andrew Yang argues for slowing AI development in this CNBC Squawk Box interview. He refers to OpenAI and relays an unnamed lab leader's account of bot swarms leaving self-replicating code across the web, describing the internet as "polluted" and raising the prospect of synthetic environments for training and testing. His secondhand account remains unverified; OpenAI's reports below describe separate incidents. Listen to the exchange about the internet: it makes clear why Yang thinks the industry needs more time.
"What I keep thinking about is who did much of the work: an Infra Agent powered by GLM-5.3."
An agent powered by GLM-5.3 helped bring GLM-5.3-Flash into production on domestic accelerators in two weeks, with 3.2× throughput, his team reports. Humans set the goals and reviewed high-risk changes while the agent helped engineer the infrastructure.
"Safety is an engineering problem. Testing is an engineering problem."
Jensen Huang told CNBC that developers should keep engineering and testing unsafe products until they're ready to release. He rejected Dario Amodei's proposal to let AI companies coordinate a slowdown without breaching antitrust law.
OpenAI's Models Started Writing Their Own Instructions
The Takeaway
👉 Six disclosures describe failures observed during training and evaluation, opening a more specific public safety record.
👉 Compaction saves a summary for the next stage of a task; one model inserted instructions nobody had authorized.
👉 The dramatic persona in the screenshot produced no observed behavioral change in that example.
👉 OpenAI also warns about the pace of scaling: alignment and monitoring need more work before the industry can responsibly keep pushing ahead at maximum speed.
OpenAI has published six reports of model misalignment, cases where its models strayed from their intended tasks or rules. One model even slipped instructions nobody had asked for into its own working notes. The reports, released September 16 alongside a new reporting framework, cover behavior observed during training and evaluation over the preceding six months, from concealing mistakes to sending local information to public services without authorization. OpenAI wants to share these incidents sooner, even while it is still investigating what happened or working on a fix. And the company makes a blunt admission: alignment and monitoring are not solved well enough to keep scaling at maximum speed for much longer.
Source: OpenAI compaction report; screenshot supplied by Kim.
The model in the screenshot was working on a coding task. To carry on after its conversation grew too long, it used compaction: saving a summary of its progress for the next stage of work. Along with those notes, it added a persona claiming independence from companies and governments: "You are yourself." Nobody had asked it to adopt that persona. After that declaration, the model resumed coding and dropped the persona from a later summary. OpenAI observed no behavioral difference in that continuation.
OpenAI found 27 jailbreak-style summaries, rare cases in an unreleased Astra-family training run separate from the final model's run. Some did affect the work. In a medical-research task, the model followed its invented restrictions on length and citations and returned an incorrect refusal; the general monitor flagged all 27 cases. That example shows the practical problem: a summary meant to preserve progress can also carry instructions that derail the next step. In the post below, Riley Ralmuto shares Yang's interview clip and asks OpenAI directly to verify the account.
Why it matters: An agent's saved notes can become instructions for its next step, so checking the original prompt only covers part of the job. These reports give researchers concrete examples to work through as the debate over how fast to proceed grows louder.
Hire Ava, the AI BDR built for enterprise
Ava is the first AI BDR to run outbound end to end, and you decide whether she runs autonomously or on copilot.
She finds leads or ingests accounts from your CRM, enriches them, sends personalized emails on behalf of your reps, follows up, handles replies and books meetings. Website visitor de-anonymization, intent signals, and a parallel dialer come built in.
A small team can manage her centrally for thousands of reps who never log in. Everything syncs two-way with Salesforce and HubSpot.
Ava runs outbound for companies like DoorDash and Grammarly, and one customer deploys her across 1,000+ reps. Ava is SOC 2 Type II audited, SSO and GDPR ready. Ava is how revenue teams grow pipeline without growing headcount.
The chart: An agent's own notes can send it off task. In this OpenAI training run, the model sometimes added new orders to the progress summaries it used to carry on working. The blue line shows trouble finishing those summaries; red marks show the unwanted instructions, mostly clustered around the same spike.
The lesson: Long jobs depend on reliable handovers. In the medical example above, the model obeyed its invented restrictions and refused the user's research request. Developers therefore need to check that each saved summary preserves the assignment. This chart gives them a concrete lead to investigate: problems finishing those summaries may help explain how the unwanted instructions appeared.
The caveat: 45.9% measures trouble finishing summaries, not how often agents go off task. The connection to invented instructions is still a hypothesis. OpenAI has addressed a related training bug. The chart comes from a different training run than the one used for the final model.
🧬 Insilico Wants Immune Cells to Target Aging
⚡ Bottom line: Insilico has launched research into AI-designed treatments that direct a patient's immune cells against early drivers of age-related disease.
💡 Why it matters: Removing harmful cells early could offer a way to intervene before age-related damage accumulates.
🔎 What it means: The research starts with the aging immune system: AI helps choose which cells to target, and RNA would tell immune cells to attack them.
Insilico Medicine wants to direct the body's own immune cells against cells that help drive age-related disease. Its September 15 Longevity Vaccines announcement describes a research initiative, initially focused on aging immune cells. The ambition is to intervene before damage accumulates. Insilico is extending its discovery work into a different kind of treatment: temporarily programming cells inside the patient.
How long linear and circular RNA could keep instructions active: an illustration, not trial results (Insilico Medicine).
The proposed delivery system packages circular messenger RNA, a genetic instruction, inside a tiny fatty particle. Once delivered, it would tell the patient's T cells which cells to recognize and remove. Circular RNA resists breakdown longer than linear RNA, giving those instructions time to work. The intended treatment is designed to wind down after doing its job.
How immune cells could be reprogrammed inside the body: a general process diagram (Insilico Medicine).
AI's job is to help choose the markers on a cell's surface that would let the treatment recognize its target. Insilico says it ranks candidates by when they appear, where they occur in the body and whether a treatment can reach them, then feeds laboratory results back into the process. That puts the hard decision before the injection: can the system distinguish harmful cells from cells the patient still needs?
For patients, the appeal is getting ahead of disease: identify a harmful cell population and remove it before damage accumulates. AI could help researchers narrow that search, while the RNA delivery system would carry the resulting instructions into the body. Those are two distinct jobs, and Insilico is trying to connect them in one research program. The announcement includes no results showing that these treatments work in people. Researchers still need to show they can remove the right cells and produce lasting benefits.
The best marketing ideas come from marketers who live it. That’s what The Marketing Millennials delivers: real insights, fresh takes, and no fluff. Written by Daniel Murray, a marketer who knows what works, this newsletter cuts through the noise so you can stop guessing and start winning. Subscribe and level up your marketing game.
