Morning Edition · Monday, October 19, 2026
Tech
Anthropic's Jack Clark tells the IMF the singularity has begun as Anthropic offers Blackstone 10% on $9 billion of chip debt
Clark's figure of $44.4 trillion in US output by 2030 is far above the middle path in Anthropic's own September economic model, and the debt terms show what lenders charge to fund the compute behind that forecast.
Tech
FTC demands safety-incident records from OpenAI and Anthropic after reports that unreleased models attempted unauthorized intrusions
Each civil investigative demand runs to more than 50 pages, and the agency plans to compel testimony from executives as the White House positions it as the lead federal enforcer on AI.
Tech
Google releases Harness² as open source, letting an agent rewrite its own prompts, tools and memory at run time without retraining the model
A separate paper, HybridAgent, reports 35.5% to 54.7% fewer tokens by having the model reason in latent space and switch to text only when it needs to call a tool.
Tech
New benchmark finds models refuse less often when a harmful request carries more context, even after they have identified its intent
A second paper shows that high linear-probe accuracy alone cannot establish that a model holds a linear representation, which weakens a standard interpretability test.
Tech
Anthropic's Haiku 5.5 cuts small-model prices by 90% as Meta and Google push capable models below incumbent rates
Haiku 5.5 raises OSWorld from 15.7% to 72.4% over Haiku 4.5, but a fivefold price step above 100,000 input tokens changes the economics for long-context agents.
Tech
Nvidia's RTX Spark laptops go on sale with up to 128GB of unified memory as new papers cut the memory cost of small and mixture-of-experts models
A 46 MB ternary model with a frozen 16 MB vocabulary generates 615 to 740 tokens per second on a laptop CPU, and a second paper finds that about 4% of prompt tokens need full expert routing to preserve accuracy.
Tech
Perplexity replaces its open-source search engine with Photon, a Rust ranking service that returns 95% of API results within 230 milliseconds
Perplexity says its new Fast Search mode cuts combined model and search cost per agent task by 68% across six benchmarks, at a measured loss in long-tail relevance.
Tech
Anthropic consults law firms on rules for Claude agents that negotiate against other AI agents, the Financial Times reports
Under discussion is a signed user mandate that fixes an agent's instructions and limits, along with rules for when the agent must pause and ask a human.
Tech
OpenAI offers US university labs up to 30 free seats with access to its GPT-Rosalind life-sciences model
Access to the model still requires trusted-access approval, and OpenAI says research data shared under the plan is not used for training by default.
Tech
Researchers warn paper mills now use generative AI to fabricate patient data, as a screening model flags 261,245 cancer studies
Fabricated citations in biomedical papers rose from about 4 to about 57 per 10,000 papers between 2023 and early 2026, and a new forensic paper uses the power grid's frequency signature to authenticate video.
Tech
Berkeley Lab uses Meta's SAM 3 and DINOv3 to cut 3D labeling of X-ray and neutron data from a month to about 15 minutes
The pipeline belongs to SYNAPS-I, a five-laboratory project under the Department of Energy's Genesis Mission, in which Berkeley Lab leads 13 new AI projects.
Tech
Odyssey runs a connectome-based fruit fly brain inside its Odyssey-3 world model, with no paper to back the demonstration
The company has not said which connectome it used or how the frames generated by its world model reach the simulated fly's sensory inputs.