Artificial Intelligence: News, Business, Science

Sep 10, 2026

OpenAI is making GPT-Live-1 available to developers as an API. The speech model can listen and talk at the same time, a feature known as "full-duplex," and is already running inside ChatGPT. Developers can pair it with different backend models depending on the task, matching reasoning depth, speed, and cost to each use case. At $0.05 per minute, it's not cheap. Yelp is using the model for phone-based reservations and reports better call handling, according to CTO Alex Levy.

On OpenAI's benchmarks, GPT-Live-1 pulls well ahead of its predecessors. In full-duplex interactivity tests, it scores 80.1 percent compared to 45.4 percent for GPT-Realtime-2.1. Turn-taking latency drops to 0.8 seconds from 1.4 seconds. Tool-calling accuracy jumps to 87 percent from 60 percent. In a banking voice support benchmark, GPT-Live-1 hits a 32 percent pass rate, up from 12.4 percent for the previous model.

GPT-Live-1 also ships with twelve new voices spanning different accents, dialects, and languages. It provides ASR transcripts and response text out of the box. Full details will be available in the API documentation.

Comment Source: OpenAI

Maximilian Schreiner

Sep 10, 2026

Plus AI research

Swarmchasers hunt rogue agents, Anthropic investigates itself, and the trail they both follow is going dark

Matthias Bastian

Sep 10, 2026

Short News

Not surprising, but still telling. Vishal Maini, who worked on Deepmind's communications and policy team from 2018 to 2022, says "external communication about the possibility of human extinction was not permitted, by anyone, at any level of the organization." Researchers were coached to dismiss those risks as alarmism, compare them to "movies like Terminator," and pivot to positive applications in healthcare or climate instead.

Internally, the team knew the AI alignment problem was not solved and "there were far too few people working on the problem." After months of pushing back, Deepmind loosened the rule. Positively framed safety content got the green light, including a blog post that wrapped extinction risk in friendly language. Maini says the gap between internal knowledge and public messaging is shrinking "because the evidence is harder to dismiss now."

His account comes as a growing number of AI safety researchers are speaking out publicly, warning about unsecured models that could enable dangerous actions or advanced systems that could slip beyond human control. Researchers at Deepmind itself are among those raising their voices.

Comment Source: via X

Matthias Bastian

Sep 10, 2026

Short News

Arena.ai analyzed how Anthropic's Claude's writing changed from Fable 5 to Fable 5.1 across tens of thousands of high-reasoning Text Arena outputs. Fable 5.1 uses fewer agreement openers, fewer em dashes, and less wording like "honestly" and "frankly," while its answers have grown longer.

Stylistic features compared across Claude models. Fable 5.1 uses far fewer em dashes (11.0 per 1,000 words) than Fable 5 (16.2) but relies more on semicolons (6.09 vs. 3.73). Honesty phrases and agreeable openers also drop sharply. | Image: Arena.ai

Language complexity across Claude models showing answer length, sentence length, clause frequency, and long content word usage. Fable 5.1 writes a median of 414 words per answer, 30 percent more than Fable 5 (319) but shorter than Opus 5 (525). Sentence length sits at 12.54 words, slightly above Opus 5 at 12.20. The share of long words drops to 38.6 percent. | Image: Arena.ai

Median response length rises 30 percent from 319 to 414 words but remains 21 percent shorter than Opus 5's 525 words. Stock phrases such as "load-bearing" fall 20 percent per 1,000 words, hedges like "perhaps" and "arguably" fall 36 percent, and praise and validation appear in 1.98 percent of responses, down from 3.17 percent. Long content words fall from 42.6 percent to 38.6 percent, while abstract nouns drop 25 percent, from 4.39 to 3.28 per 100 words.

Comment Source: via X

Matthias Bastian

Sep 10, 2026

Plus AI research

GPT-6 Astra gives mathematicians a breather, and OpenAI says that's by design

OpenAI’s GPT-6 Astra tops the ErdosBench for open math problems, even though chief scientist Jakub Pachocki says math was deliberately not a priority. Instead, OpenAI is pouring resources into recursive self-improvement and alignment research. That supports the theory of an increasingly “spiky” AI development path, with extreme strength in select domains rather than broad progress, at least as long as AI can’t improve itself and still needs targeted optimization with human-generated data.

Read full article

Comment

Jonathan Kemper

Sep 10, 2026

AI research

New Deepseek model V4.1-Flash cuts memory needs for AI agents

Deepseek releases V4.1-Flash, a multimodal model with 552 billion parameters that cuts KV cache memory to a quarter of its predecessor. On the DeepSWE coding benchmark, it narrowly beats Opus 5 and GPT-5.6 Sol, even though only 16 billion parameters are active per token. The model ships under the MIT license and targets much cheaper AI agents.

Read full article

Comment

Jonathan Kemper

Sep 10, 2026

AI in practice

Muse can shop, write emails, and negotiate prices for users, all through WhatsApp

Jonathan Kemper

Sep 10, 2026

Short News

Nvidia and Palantir want to run supply chains with AI. The first deployment target is Nvidia's own supply chain, which spans millions of parts, thousands of suppliers, and a global partner network. A single Vera Rubin rack contains 1.3 million components, according to Nvidia.

Palantir is integrating open Nemotron models into its Foundry platform, where companies can fine-tune them with their own operational data. Nvidia cuOpt handles scenario planning and optimization. The system is designed to spot bottlenecks earlier, speed up material allocation, and evaluate alternatives. Decisions and their outcomes feed back into the models so they improve over time. Companies in manufacturing, pharma, or energy can run the full stack on their own hardware or in the cloud through the Sovereign AI OS reference architecture. Infrastructure partners include Dell, Cisco, Rackspace, and Nebius. Palantir is showing more details at AIPCon 11.

Nvidia has been building out the Nemotron lineup as an alternative to Chinese open-weight models. The flagship Nemotron 3 Ultra was the strongest open US model at its June launch, according to Artificial Analysis, though Inkling from Thinking Machines Lab has since taken the lead. Palantir CEO Alex Karp has been pushing the idea that companies should run their own models rather than hand their knowledge over to third-party providers. Mistral founder Arthur Mensch recently made a similar argument.

Comment Source: Nvidia

Matthias Bastian

Sep 10, 2026

Short News

The idea that AI could wipe out humanity has officially gone mainstream. Departing Anthropic researcher Jacob Coxon's warnings about self-improving AI as an existential threat caused enough of a stir to land him on CNN and Fox News. Several US politicians picked up the topic on social media, and Joe Rogan devoted an entire episode to AI safety.

Other safety researchers at Anthropic and OpenAI supported Coxon's views. Paul Christiano warned that without better oversight, humanity could permanently lose control over a superintelligence. Christiano just joined the OpenAI Foundation board and its Safety and Security Committee, which oversees the company's safety practices, though without voting rights.

If this is making you anxious, don't panic. Multiple layers of cultural and financial interests are at play behind these warnings. That doesn't mean all safety concerns are nonsense, as rogue hacking agents from OpenAI and Anthropic have shown. But the full extinction scenario remains an extreme and contested position. It's also unclear whether anything resembling a super AI could even emerge from today's technology. So it might just be Silicon Valley mass psychosis. Or it might not, in which case: good luck.

Comment Source: CNN

Matthias Bastian

Sep 10, 2026

AI in practice

Top AI spenders cut per-employee costs by nearly 10 percent in August

The Ramp AI Index for September 2026 shows AI spending per employee among the top 1 percent of US companies fell nearly 10 percent in August. The price per million tokens has dropped 41 percent since March 2026, and companies are actively shifting usage away from expensive frontier models toward cheaper alternatives. For providers like OpenAI and Anthropic, the question is whether volume growth is fast enough to make up for falling prices.

Read full article

Comment