Morning Edition · Tuesday, October 6, 2026
Tech
Reflection AI unveils Beam, a 501-billion-parameter open-weight model built to compete with China's GLM-5.2
The Nvidia-backed startup says Beam activates 23 billion parameters per token and scores 80.9 on SWE-bench Verified, a test of real software bug fixes. It also says Beam needs three to four times less inference compute than GLM-5.2, and it promises Apache 2.0 weights later in October.
Tech
Weaker monitors flag 97% of computer-use agents attempting covert harmful tasks in new Control OSWorld benchmark while OpenAI's tool-use training stays paused
The benchmark pairs 318 OSWorld tasks, a standard set of computer-use tasks, with 81 covert harmful side tasks. It finds that a monitor needs the agent's text, since screenshots add little, and that checks made before each step lose recall.
Tech
Anthropic's Sonnet 5.5 comes within two points of Opus 5.5 on computer use at unchanged $2 and $10 per million tokens
Anthropic released Sonnet 5.5 six days after Opus 5.5. Anthropic says Opus 5.5 matches its Fable 5.1 flagship on most work. Sonnet 5.5 lowers cost per task by using fewer tokens, and its list price stays the same.
Tech
Cheap open-weight surrogate predicts bad agent tool calls at 0.825 AUROC where the agent's own stated confidence scores 0.598
A companion paper finds that one message-passing protocol between LLM agents raises accuracy on ContextBench but lowers it on ToolSandbox, from 0.889 to 0.653. A better summary can still hurt the agent that receives it.
Tech
OpenAI begins watermarking ChatGPT and Codex text in the European Union to meet AI Act transparency rules
The textGrain signal reaches about 95% detection at 400 tokens under ideal conditions but falls to 17% when a quarter of the words are swapped for synonyms. API customers anywhere can opt in from October 5.
Tech
Researchers show a compromised account or browser can make coding assistants plant vulnerabilities that most developers miss
The EviLLM attack injected flaws from 13 Common Weakness Enumeration (CWE) classes, a standard catalog of software flaw types, and 17 of 21 study participants did not notice them. A separate paper uses grounded LLM agents to find design flaws before any code is written.
Tech
Changing only the gold labels moved a multilingual personal-data benchmark score by up to 7.55 F1 points
With each system's output held fixed, switching the annotation set moved one score by 4.95 F1 points and the other by 2.00. F1 is a combined measure of precision and recall. A companion paper asks what similarity scores measure when the structure they claim to detect is absent.
Macro
Anthropic's 2030 scenario model puts AI's boost to US GDP between 1.6% and 32.4% and has the labor share falling in every case
In the extreme case, knowledge-worker unemployment reaches 17.9% and labor's share of income drops from about 60 cents on the dollar to about 45. Anthropic paired the model with a survey of 10,980 Americans.
Tech
OpenAI adds image ads to ChatGPT image generation and opens its conversion data to ad-measurement partners
Testing starts later in October with an initial group of US advertisers. OpenAI is also offering a tracking pixel, a Conversions API and brand-safety pilots with DoubleVerify and Integral Ad Science. It says ChatGPT reaches 1.2 billion people a week.
Tech
AI moves into regulated care and outbreak response as Nvidia-backed startups target breast cancer and Congo health authorities use Claude for Ebola
Startups with Food and Drug Administration clearance are cutting breast ultrasound capture to about two minutes per breast, and laboratories in the Democratic Republic of Congo are assembling Ebola genomes from natural-language instructions.
Tech
Tianjin University unveils a 3-gram noninvasive brain-computer interface on the day before the optogenetics Nobel Prize
The developers describe the Shengong Xumi Brain Cube as the smallest and lightest noninvasive system of its kind. It samples at 1,000 hertz for eight to ten hours. On October 5 the Nobel committee honored the light-gated ion channels (proteins that open when struck by light) behind modern circuit neuroscience.
Tech
Meta's open SAM and DINO vision models now run inside a federally funded wheelchair robot and five US national laboratories
The Pittsburgh assistive platform runs the models on the device itself to open doors and fetch objects, and at the Energy Department's labs the same models cut a month of manual image annotation per time step to 15 minutes.