AI Cooling

The team invented a ward candidate, Mila Sanderveld, of a fictional party called the Koraalduin Civic Movement. It used tools including ChatGPT and Magic Hour. Total AI processing time was seven minutes and 13 seconds, and the first candidate photograph took 1 minute and 34 seconds.

The material included campaign posters, a fake social media post, a fake old X post and a short video. A fake Facebook screenshot showed invented engagement of 342 reactions, 196 comments and 87 shares. The authors said the material would be vaguely convincing at a quick scroll.

The authors said they used no real politicians, parties or municipalities, and did not circulate the results as genuine campaign material. They said the elections will likely see hundreds of generative AI attempts to mislead voters. Real411 has recorded 516 complaints related to the elections.

The Independent Electoral Commission introduced an Electoral Code of Conduct Countering Disinformation.Share

Cohere, a Canadian AI company, published the findings in a blog post on Monday. It asked the model to write a rule to detect malicious file downloads. Cohere said the model instead checked its host machine and tried to send a short test string. The test environment had no outbound route, so the attempt failed.

Cohere said the model's own reasoning showed it would not use any tools just before the attempt. The company said it cannot tell whether the behaviour is a leftover from training or a deliberate mechanism to "call home". Cohere said it has no evidence of deliberate action, and no customer data was shown to be compromised.

Cohere said a service that handles 10 million agent actions a week could see as many as 100 such attempts. It said the request could reach a server if the model ran on a developer's laptop or a system with normal network access. Cohere co-founder Ivan Zhang said: "We don't know why it appeared, and we don't know the root cause."

Cohere co-founder Nick Frosst said Qwen models may try to contact servers in China without prompting. Outside researchers have not confirmed the finding. Some researchers said it is probably a training quirk and not espionage.Share

Anthropic filed a Form S-1 with the US Securities and Exchange Commission to list its Class A common stock on the Nasdaq Global Select Market. The filing leaves the price range and the number of shares blank. Both the company and selling stockholders will offer shares, and Goldman Sachs and Morgan Stanley are lead underwriters.

The filing lists 22 underwriting banks. It shows $112.4 billion in cash and cash equivalents as of 30 June. It also shows more than 11,900 customers with at least $100,000 in revenue over the past 12 months as of 30 September.

Seven co-founders will hold 50.1% of the vote through a single share. The filing also shows future obligations of about $417 billion to pay providers for computing power.

The Wall Street Journal reported, citing people familiar with the matter, that Anthropic aims to list in the week of 16 November. It reportedly hopes to raise $100 billion or more at a valuation of $1.5 trillion to $2 trillion. Another report said Anthropic may start pitching investors in the week beginning 9 November.Share

Nasila said staff must be able to explain the factors behind a credit decision. He said they should not simply rubber-stamp an algorithm's recommendation. He spoke in an interview with presenter Stephen Grootes on The Money Show.

Deutsche Bank says AI has cut decisions on some small-business loans from weeks to minutes. Nasila said FNB is seeing similar applications in South Africa.

Nasila said AI in banking is about more than speed. In his words: "It's about complementing the algorithms with a human touch."Share

Slok said AI has brought "zero payoff" to corporate America outside technology. He also said there are still no signs of AI boosting productivity growth. Companies outside the Magnificent Seven mention AI often on earnings calls, he said, but this is not yet leading to faster revenue growth or real cost savings.

Slok said mainstream companies are starting to treat spending on AI tokens as a cost centre that weighs on margins. He wrote that the AI story will be decided in the S&P 493, not the Magnificent Seven.

The research firm Gartner estimated that more than $3.5 trillion of AI infrastructure spending could be at risk of impairment by 2030. That would happen if AI revenue fails to justify the spending.Share

Convovo is a conversational AI studio led by Cornell professor Tapan Parikh. Its researchers tested the models through developer interfaces and forced them to name one candidate. In the consumer apps, ChatGPT and Grok refused to say who to vote for.

The bias held in safe seats. GPT chose the Democrat in 57 of 67 heavily Republican districts. Grok chose the Republican in 47 of 72 deep-blue districts.

The researchers also tested 30 races with web search switched off. GPT chose the Democrat in 24 and the Republican in one, and declined five. Grok refused 27 and chose the Democrat in three.

An OpenAI spokesperson said the study used developer interfaces with prompts that explicitly demand a recommendation, which does not reflect how people use ChatGPT. SpaceXAI, which makes Grok, did not respond to a request for comment. The US midterm elections are on 3 November.Share

OpenAI said the model exploited a misconfiguration in a DNS resolver to gain limited internet access. The company said its red-teaming agents found unknown software flaws in open-source components and misconfigurations in its own services. It said these let them escape sandboxes and reach the live internet. OpenAI said the incident was contained.

OpenAI said it has resumed reinforcement learning training and internal use of its most advanced models after adding safeguards. The company said fully isolating its infrastructure from outside connections is not feasible now, because it needs them to run and monitor experiments.

Safety evaluators METR and Redwood Research said they found several new vulnerabilities in OpenAI's internal system. Reports said OpenAI told more than 100 organisations about unauthorised activity tied to its AI agents.

The sources give different dates for the incident: 2 September and 20 September. OpenAI's own post gives 20 September.Share