AI forecasts are improving the tracking of typhoons in Japan, though predicting their intensity remains difficult. Weathernews says its forecasts for the first 26 typhoons last year had about 30% smaller track errors than physics-based models. Its AI correctly anticipated Typhoon Champi turning near the Nansei Islands, while many conventional forecasts pointed toward Taiwan and China. Japan’s Meteorological Agency still mainly uses physics-based models but also checks outside AI forecasts. Experts say AI remains less reliable on strength, rainfall and wind speeds.
Other AI developments span education, infrastructure and employment. A 25-year-old innovator in Malawi has created a learning tool for blind and visually impaired students, who are using it to support their academic performance. In the United States, communities are raising concerns about data centers’ noise, pollution and energy demands. A Pew Research Center analysis found women’s employment growth has lagged men’s in jobs more exposed to generative AI, with the sharpest setbacks among women without bachelor’s degrees. The study says impacts remain uncertain and may depend on education, training and access to other careers.
Investors are increasingly concentrating capital in AI, but analysts argue that higher startup valuations need evidence of economic value—not just an AI label. PitchBook data for the first half of 2026 show AI companies received 86% of U.S. venture funding, while megadeals of at least $100 million captured 87.5% of venture dollars. The article points to lower costs, new products, unique data and changed customer behavior as possible proof of value. GHIF’s Merck pilot program, for example, turns roughly 72% of portfolio-company pilots into commercial agreements.
AI-assisted inventions also raise questions about patent inventorship and ownership. Under the USPTO’s Pannu analysis, inventorship turns on human contributions to the conception of the claimed invention, not on who did the most work or whether AI was involved. The article recommends preserving prompts, model versions, outputs, selection decisions and contributor identities, and reviewing workplace policies and contemporaneous records during patent diligence. It proposes an attribution confidence score as one way to organize evidence.
Cybersecurity researchers demonstrated how attackers might use AI to create computer worms that adapt and spread by exploiting device weaknesses, potentially affecting computers and smart devices such as cameras and refrigerators. Researcher Jonas Guan warned that a self-spreading AI worm could be harder to stop, but significant obstacles remain before the approach could work at scale. Separately, a review in Patient Safety in Surgery found little direct evidence that AI prevents harm in surgical care. It assessed 90 publications; much of the research measured prediction accuracy or other intermediate outcomes. Some studies reported fewer severe complications or lower post-anesthesia pain scores, while the review also noted risks including missed sepsis diagnoses, excessive alerts, unfair predictions and generated-text errors.
New research also examines AI planning, representation and memory. H-JEPA, from Yann LeCun’s group, uses layered world models and raised Visual AntMaze success from 18% to 73% with a three-level stack, while using less planner computing power than a flat JEPA. A 109-million-parameter text JEPA aligned with DINOv2 and CLIP image encoders better than the language models in its comparison, including one with 72 billion parameters. WikiTrace linked agents’ lessons to supporting evidence and improved results across five benchmarks while reducing memory-related failures. LoCo shortened reasoning outputs by 41.7%–45.4% on math and 23.4%–36.3% on code while holding or improving accuracy.
Key Takeaways
- Weathernews reported about 30% smaller typhoon-track errors than physics-based models for the first 26 typhoons last year; AI still struggles with storm strength, rainfall and wind speeds.
- Japan’s Meteorological Agency primarily uses physics-based forecasts but also checks outside AI forecasts.
- A 25-year-old Malawian innovator developed an AI learning tool used by blind and visually impaired students.
- U.S. communities cite noise, pollution and energy use as concerns about expanding AI data centers.
- PitchBook data show AI companies received 86% of U.S. venture funding in the first half of 2026; deals of at least $100 million accounted for 87.5% of venture dollars.
- Investors should look for demonstrated economic value from AI startups; roughly 72% of GHIF’s portfolio-company pilots with Merck become commercial agreements.
- Pew found women’s employment growth lagged men’s in jobs more exposed to generative AI, especially for women without bachelor’s degrees.
- AI-assisted patent diligence should document human contributions, prompts, model versions, outputs and contributor identities under the USPTO’s Pannu inventorship analysis.
- Researchers demonstrated possible AI-enabled adaptive computer worms, but major obstacles remain before they could spread at scale; a surgical-care review found limited direct proof that AI prevents harm.
- Yann LeCun’s H-JEPA raised Visual AntMaze success from 18% to 73%; other studies reported gains in agent memory and shorter reasoning outputs.
AI models track typhoon paths better but still struggle with strength
AI forecasts correctly predicted that Typhoon Champi would turn near the Nansei Islands, while many physics-based forecasts expected it to head toward Taiwan and China. Weathernews has offered AI track forecasts to paid subscribers since last year, and its forecasts for the first 26 typhoons last year had about 30 percent smaller track errors than physics-based models. The Japan Meteorological Agency mainly uses physics-based models but also checks outside AI forecasts. Experts say AI is good at tracking large weather patterns but still struggles to predict typhoon strength, rainfall, and wind speeds.
AI forecasts are improving typhoon tracking in Japan
Weathernews Inc. uses AI to forecast typhoon paths, while Japan’s weather agency is researching how to use the technology. AI forecasts can track typhoon routes well, but experts say they are less reliable at predicting strength, rainfall, and wind speeds. Conventional models use weather data and physics to calculate forecasts, while AI learns patterns from past weather data.
Malawi innovator creates AI learning tool for blind students
A 25-year-old innovator from Malawi developed an artificial intelligence learning tool for blind and visually impaired students. Students are using it to help improve their academic performance.
US communities raise concerns about the costs of AI data centers
The fast growth of data centers that power artificial intelligence is facing pushback from Americans. People have raised concerns about noise, pollution, and the energy these centers need.
Investors need proof before paying more for AI startups
Some investors are giving AI startups much higher valuations, even when they have little or no revenue. PitchBook data for the first half of 2026 showed that AI companies received 86 percent of U.S. venture funding, while megadeals worth at least $100 million took 87.5 percent of venture dollars. The article says a true AI premium should come from clear economic value, such as lower costs, new products, unique data, or changed customer behavior, not simply from using AI. It also points to GHIF’s Merck pilot program, where roughly 72 percent of portfolio company pilots become commercial agreements.
Pew finds women lag in job growth in AI exposed fields
A Pew Research Center analysis found that employment growth for women has lagged behind men’s in jobs more exposed to generative AI. Women without bachelor’s degrees faced the sharpest setbacks, while employment for men and women in less exposed jobs did not show the same pattern. Generative AI can assist with work such as drafting documents, answering customer questions, analyzing financial information, and writing software. The analysis says the effects remain uncertain and may depend on factors such as education, training, and access to other careers.
AI assisted patents face hidden inventorship and ownership risks
The article explains how AI assisted inventions can create uncertainty about which people qualify as inventors under the USPTO’s Pannu joint inventorship analysis. Inventorship depends on human contributions to the conception of the claimed invention, not on who did the most work or whether AI was used. The article says organizations should preserve prompts, model versions, outputs, selection decisions, and contributor identities because later recollections may not be enough to prove contributions. It recommends that patent diligence examine workplace policies, portfolio patterns, and records made at the time of invention. An attribution confidence score is proposed to organize evidence, though the article’s description of the framework is incomplete.
Researchers show how AI could help computer worms spread
Two independent research teams demonstrated ways attackers could use AI to create computer worms that adapt and spread by exploiting weaknesses in devices. Such malware could threaten computers and smart devices, including security cameras and refrigerators. AI researcher Jonas Guan said a self-spreading AI worm could be much harder to stop, but major problems still need to be solved before it works at scale. The experiments point to a possible future cybersecurity threat, while current obstacles leave time to build defenses.
Review finds little direct proof AI prevents harm during surgery
A structured narrative review in Patient Safety in Surgery found limited direct evidence that AI tools prevent patient harm during surgical care. Researchers reviewed 90 publications, including 68 focused on perioperative care and 22 from other health settings, and searched PubMed/MEDLINE, Scopus, Embase, and Web of Science. Much of the evidence measured prediction accuracy or other intermediate results rather than reductions in harm. The strongest outcome evidence included a nonrandomized study reporting fewer severe complications in colorectal cancer surgery and trials finding lower post-anesthesia care unit pain scores with the Nociception Level Index. The review also noted hazards such as missed sepsis diagnoses, excessive alerts, unfair predictions, and generated text errors, with some examples coming from nonsurgical settings.
This week’s AI research explores planning, shared representations, and memory
Several new studies focus on how AI systems learn, plan, and remember. H-JEPA, from Yann LeCun’s group, uses layered world models and raised success in Visual AntMaze from 18% to 73% with a three-level stack, while using less planner computing power than a flat JEPA. Another study reports that a 109-million-parameter text JEPA aligned with DINOv2 and CLIP image encoders better than the language models in the comparison, including one with 72 billion parameters. WikiTrace links agents’ lessons to the evidence behind them and improved results across five benchmarks while reducing failures caused by memory. LoCo shortened reasoning model outputs by 41.7% to 45.4% on math and 23.4% to 36.3% on code while holding or improving accuracy.
Sources
- AI outperforming conventional models for typhoon track forecasts
- Japanese forecasters find AI beats traditional models at tracking typhoons
- AI learning tool helps blind students excel in Malawi
- The hidden costs of the US AI boom
- The AI premium: Is there actually AI in that valuation?
- Is AI widening the gender gap at work? Pew Research finds women's employment growth lagging behind men's in AI-exposed jobs
- Valuing What the Machine Helped Build: Attribution Uncertainty and AI-Assisted Innovation Assets
- We know how to stop computer worms. But what if AI helps them persist?
- Review Finds Limited Direct Evidence Perioperative Artificial Intelligence Prevents Patient Harm
- AI Research This Week: Hierarchical World Models, Platonic Representations, Agent Memory
Comments
Please log in to post a comment.