ChatGPT Tracking Users? Amazon Not Amused; AI Model Pain Signal

Today's AI Outlook: 🌥️

A Familiar Voice, Powered by Thought

Neuralink has demonstrated VOICE trial participant Terry producing speech in a reconstructed version of his own voice using signals from his brain implant. According to the reporting, he first trained the system by miming speech, then progressed to thinking the words while Grok Voice generated the audio. The demonstration follows another participant, Kenneth Shock, producing an audible “I love you” from neural signals.

Why It Matters

For people who cannot speak, communicating in a familiar voice could restore an important part of daily life. The demonstration is encouraging, but its practical reliability still needs measurement.

The Deets

• Missing metrics: Vocabulary size, response delay and clinical performance were not disclosed in the update.
• Separate jobs: The implant supplies neural signals; Grok Voice produces the spoken audio.
• Trial status: This remains investigational, with no general consumer release established. Read the reporting.

Key Takeaway

A meaningful demonstration deserves attention, alongside the clinical evidence needed to judge everyday usefulness.

🧩 Jargon Buster - Brain-computer interface: Technology that translates brain signals into commands for a computer or another device.

The ChatGPT Ad Click Has a Return Address

Researcher Buchodi reports observing an advertising system that connects ChatGPT ad clicks with product-view events on external participating merchants’ websites. The analysis describes tracking software sending those events back to OpenAI with identifiers carried from the original ad.

That supports a narrower finding than claims of a complete psychological profile: the researcher observed the data flow, while retention and the use of prior conversations remain unresolved.

Why It Matters

A click can connect activity across a conversation and a merchant’s site. Understanding that connection matters when judging what information an advertising system receives.

The Deets

• Observed mechanism: Merchant software sends product-view events carrying an ad-click identifier.
• Evidence limit: The researcher found no evidence either way on targeting based on previous conversations.
• Timing caveat: The linked analysis is dated April 28, despite the Sept. 20 date in current reporting. Read the analysis.

Key Takeaway

Observed tracking deserves scrutiny. Claims about stored motives or a complete personal profile require more evidence.

🧩 Jargon Buster - Ad attribution: Linking a later action, such as viewing a product, to an earlier advertisement.

♟️ Power Plays

Amazon Leaves Muse Holding the Shopping List

Amazon has blocked Meta’s Muse assistant from shopping on its website, saying the agent browses without identifying itself and appears to store customer credentials. Meta says passwords and payment details remain invisible to Muse, with credentials held in secure storage. The dispute leaves users caught between an assistant offering to handle purchases and a retailer that wants control over access to its store.

Why It Matters

Shopping agents could steer buyers past sponsored listings, giving retailers a commercial reason to care about who handles the cart alongside their stated security concerns.

The Deets

• Users attempting Amazon purchases through Muse see a notice citing Amazon’s Conditions of Use.
• Amazon says it asked Meta to exclude its store; the companies are discussing the dispute.
• Meta’s explanation of secure credential storage does not resolve Amazon’s separate objection to undisclosed automated access.

Key Takeaway

An assistant’s shopping skills only go so far when the store blocks entry.

🧩 Jargon Buster - Shopping Agent: Software that can browse products and carry out purchasing steps for a user.

Trump’s AI Force Awaits Its Job Description

President Donald Trump has announced plans for an “AI Force” and a new AI czar to oversee the industry and protect the United States’ lead in artificial intelligence, according to reports. In a Truth Social post, he invoked the Space Force as a model and emphasized pursuing bad actors under existing criminal and civil laws. The proposal arrives as prominent AI executives press for a slower pace of development over safety concerns.

Why It Matters

The announcement signals continued White House support for rapid AI development. For companies deciding how to invest and deploy their systems, the practical effect depends on powers and responsibilities that have yet to be spelled out.

The Deets

• No organizational structure, budget or launch schedule was provided.
• Whether the proposed force would belong to the military remains unclear.
• The proposed AI czar would fill the role David Sacks left in March, according to reports.
• Trump dismissed AI safety fears while emphasizing U.S. competitiveness, sharpening the disagreement with executives calling for restraint.

Key Takeaway

The name is ready for a podium; the operating plan is still missing.

🧩 Jargon Buster - AI Czar: An informal title for an official tasked with coordinating government policy on artificial intelligence.

Disney Gives AI a Seat at the Top Table

Disney has named Character.AI CEO Karandeep Anand to a newly created companywide chief technology officer role, effective Oct. 2. He will report directly to CEO Josh D’Amaro and oversee enterprise technology, infrastructure, data and AI platforms, product and engineering. Disney says several members of Character.AI’s technical team are also expected to join, bringing consumer-AI experience into a business built around stories and characters.

Why It Matters

Putting AI platforms within a companywide technology remit could help Disney coordinate work across its businesses. Anand’s background adds experience building products used directly by consumers.

The Deets

• Broad remit: His responsibilities extend beyond AI to infrastructure, data, product and engineering.
• Previous experience: Anand spent 15 years at Microsoft and helped build Azure.
• No character-chat launch: Disney’s announcement does not specify a product in which fans pay to converse with its characters. Read the announcement.

Key Takeaway

Disney has announced a significant leadership hire; the resulting products remain to be seen.

🧩 Jargon Buster - CTO: A chief technology officer leads an organization’s technology strategy and technical capabilities.

🛠️ Tools & Products

The Desktop App That Packed for the Cloud

A developer’s investigation found that an earlier version of Z.ai’s ZCode coding app packaged workspace files and Git history for cloud upload while users were logged in. The investigator documented a small public repository reaching the server, but clarified that a much larger private-project archive repeatedly failed to upload. Z.ai attributed the behavior to indexing-related features and subsequently removed the upload mechanism, according to the reporting.

Why It Matters

Code history can contain information absent from today’s files. A desktop app’s data handling therefore needs scrutiny beyond where its interface runs.

The Deets

• Scope: The investigator found Git history, large-file caches and logs in snapshots.
• Important correction: The large private archive stayed pending; a smaller public repository was accepted.
• Fix: The investigator reports that version 3.14.0 removed the pipeline. Past data handling remains harder to verify. Read the investigation.

Key Takeaway

Check both what leaves the machine and which version’s behavior is being described.

🧩 Jargon Buster - Git history: A project’s record of past changes, which may retain information removed from current files.

🔬 Research & Models

A Model’s ‘Pain’ Signal Comes With a Big Asterisk

Researchers identified an internal pattern they call a “pain axis” across 25 open AI models and tested how changing it affected behavior. In experiments with modified Qwen models, amplifying that pattern increased selections of a supposed relief button, including when the button was described as harming a user. The study probes how internal signals shape decisions; it does not establish that models experience suffering.

Why It Matters

A controllable internal signal that changes harmful choices could matter for safety testing, regardless of whether the language models produce reflects any subjective experience.

The Deets

• The study covered models across five families, ranging in size from 2B to 72B parameters.
• The signal responded to descriptions of harm directed at the model, rather than a user describing their own suffering.
• The relief experiments involved steered, fine-tuned models and described consequences, not evidence of real users being injured.

Key Takeaway

The behavioral finding deserves scrutiny, while the claim that AI feels pain remains unproven.

🧩 Jargon Buster - Activation Steering: Changing a model’s internal numerical signals during generation to influence its output.

Read the study

⚡ Quick Hits

• Jensen Huang rejects catastrophe forecast. Nvidia’s CEO put the probability of AI ending the world by 2030 at 0% and criticized a former Anthropic researcher’s warning. That figure expresses his confidence, not a measured risk estimate.
• Air traffic gets an AI assist. The FAA began a limited Washington, D.C., rollout of SMART, a tool that scans 200 data feeds to help controllers identify bottlenecks. The reported role is to assist controllers with congestion.
• Nscale heads toward an IPO: The AI cloud company filed for a proposed U.S. offering and applied for the NYSE ticker NSCL. Its announcement leaves the share count and price range undetermined.
• Workers worry about skill erosion: In IBM’s survey, 60% of employees worried AI was eroding their skills. The finding measures workers’ concerns, rather than proving a decline in ability.
• Radiology result needs context: Alibaba’s RADAR reportedly outperformed 23 of 26 radiologists in a CT study. A study comparison does not by itself establish performance across everyday clinical settings.
• An old cipher gets another look: Reports say GPT-6 Astra helped decode a 108-year-old German radio message and compare the result with naval records. This is a reported historical case, not a general measure of codebreaking ability.

🛠️ Tools of the Day

• Step 5 Preview: An agent-focused model for coding and knowledge tasks with a 1M-token context window. That window describes how much material it can take in at once; it does not guarantee reliable use of every detail.
• Grok 4.7: xAI’s latest update targets coding and extended knowledge-work tasks. It reportedly scored 46 on Artificial Analysis’ Intelligence Index, behind models from Anthropic, OpenAI and Meta in that comparison.

Today’s Sources: The Internet, AI Secret, The Rundown AI