Astra Outperforms OpenAI's Sol in Real-World Lab Data Test

Key Takeaways

• Astra outperformed OpenAI's Sol models in messy lab data tests, reaching 66.7 percent accuracy versus Sol at 62.0 percent
• Employees who used ChatGPT at work were trusted at only 24 percent compared to 68 percent for non-users, even at AI-friendly companies
• Microsoft, Barclays Bank, and Kroger are rehiring workers after AI cuts failed to deliver expected results
• Mito produced over 5000 AI assets for a World Cup TV ad using Google Cloud and models like Omni, Veo, and Nano Banana
• Nearly half of severely distressed Israeli users seek emotional support from ChatGPT or Gemini, though researchers warn AI lacks real empathy
• Israeli researchers created an AI self-teaching method without human data, but models cannot learn real-world facts without real data
• Cardinal Michael Czerny called for communal dialogue on AI safeguards at the Builders AI Forum in Rome
• Nanjing University rolled back a smart classroom AI system that tracked student attentiveness after public backlash
• Nvidia and Lam Research stocks fell after midterm elections due to fears of slower data center construction
• Fabian Theis says virtual cells are not ready and urges focus on harder benchmarks and real organoids

Messy lab data reveals true AI limits

A new benchmark tested AI models on real wet-lab tasks with messy data instead of clean charts. Astra outperformed OpenAI's GPT-5.6 Sol and GPT-6.1 Sol in control handling and restraint. Sol accepted empty control plates and miscounted colonies by about 12 percent. Astra reached 66.7 percent accuracy compared to Sol at 62.0 percent and GPT-6 Luna at 38.3 percent. The study argues that lab agents should be tested on raw plates and failed controls before claiming reliability.

Distressed Israelis turn to ChatGPT for support

A study by Tel Hai Academic College and Bar-Ilan University found that nearly half of severely distressed Israeli users seek emotional support from generative AI tools like ChatGPT and Gemini. About 29.3 percent of all AI users surveyed said they used AI at least once for emotional help. The study was published in Cyberpsychology, Behavior, and Social Networking. Younger people and women were more likely than older people and men to use AI for mental health support. Researchers warn that AI lacks real empathy and crisis intervention skills.

Companies rehire workers after AI cuts fail

Companies like Microsoft, Barclays Bank, and Kroger have cut jobs citing artificial intelligence this year. Some of these companies are now rehiring laid-off workers because AI tools did not perform as expected. This trend has been nicknamed the AI boomerang. The reversals are happening often enough to draw attention, though it is still unclear how common they are. The phenomenon affects tens of thousands of workers across the economy.

Chinese university ends AI student monitoring

Nanjing University in Jiangsu province stopped using a smart classroom feature that tracked student attentiveness after public backlash. The system monitored late arrivals, attendance, and a head-up rate to measure focus. Leaked screenshots showed live classroom footage paired with real-time AI tracking. The images went viral and sparked debate over privacy and algorithmic surveillance. The university announced the rollback after intense criticism from students and the public.

Mito produced 5000 AI assets for World Cup ad

Mito built a serverless AI video production system on Google Cloud for a 120-second Movistar Plus+ national TV spot during the World Cup. The campaign generated over 5000 assets including 1000 videos across about 65 iteration rounds. Mito used Google Cloud Pub/Sub, Cloud Run, Cloud Storage, and Vertex AI with models like Omni, Veo, and Nano Banana. The system handled long video renders in the background so creators could keep working. The final campaign was described as 95 percent AI-generated and aired in prime time.

Virtual cells still promise more than they prove

Fabian Theis says virtual cells are not ready to deliver on their big promises. Cell data has grown to 10 million transcriptomes, but perturbation data remains very limited. Current models lack agreed ways to represent connected organoids and miss important context like immune cells. Theis is developing Lab Compass, an inverse model that uses reinforcement learning to find interventions reaching target states. He urges researchers to focus on surviving harder benchmarks and real organoids rather than virtual cells alone.

Cardinal warns AI impacts everyone but people are not listening

Cardinal Michael Czerny spoke at the Builders AI Forum in Rome about the need for collaboration on AI safeguards. He said everyone is affected by AI but people are not listening to each other. He stressed that learning about AI should be a communal dialogue, not an isolated process. The two-day summit at the Pontifical Gregorian University brought together Catholic AI companies, investors, and researchers. The goal was to build a community supporting AI that serves the Church's mission.

Israeli researchers develop AI self-teaching method without human data

Israeli researchers created a method that lets AI teach itself using only data generated by other AI models. Two AI models trained together, with one creating challenges and the other solving them, without using external sources like documents or images. The model improved its predictive skills across language, images, audio, code, and DNA sequences as computing resources increased. However, the models could not learn factual information about real-world subjects without real data. The research was published on arXiv and aligns with Israel's push to become a leader in artificial intelligence.

Midterm elections shook up Wall Street and key industries

The Democratic victory in the midterm elections caused major shifts in Wall Street stock prices. AI-related stocks like Nvidia and Lam Research fell due to fears of slower data center construction under a Democratic Congress. Defense companies including Northrop Grumman and Lockheed Martin dropped 4% to 12% on hopes for an end to the U.S.-Iran war. Oil companies like Exxon Mobil lost ground while hospital shares rose on expectations of restored government funding. Liquor companies and renewable energy firms also gained, while Elon Musk's companies SpaceX and Tesla fell after his heavy spending on Republican candidates.

Bosses forcing workers to use AI are realizing their mistake

Many companies pushed employees to use AI tools like ChatGPT to boost productivity, but bosses are now regretting that decision. A study by Samuel Meyers at Hebrew University of Jerusalem found that employees who used AI were trusted much less, with only a 24% trust rate compared to 68% for those who did not use AI. This happened even at companies that encouraged AI use. Researchers found that AI tools have also increased costs for companies. Employees now face a dilemma between following employer instructions to use AI and earning trust by avoiding it, which could lead to hidden or shadow AI use at work.

Sources

• Messy bench data is the real AI test
• Israelis in emotional distress are turning to ChatGPT for help, study finds
• What to consider if a company that made AI-related job cuts wants you back
• Chinese university stops using AI to spot absent-minded students after outcry
• How Mito orchestrated 5,000+ AI assets for a World Cup TV spot
• Virtual cells still promise more than they prove
• ‘We are all impacted by AI but nobody’s listening’ to one another, cardinal warns at conference
• Israeli researchers develop method letting AI teach itself without human-generated data
• From high-flying AI stocks to tariff-hit alcohol companies, the midterms shook up Wall Street
• Remember When Bosses Were Forcing Their Workers to Use AI? They're Now Realizing They Made a Huge Mistake