California-based Alef Aeronautics released a video on Thursday showing its electric flying car soaring over a Tesla Roadster, welcoming the automaker to what it calls the “flying car race.”
The company said the demonstration marked its 1,000th logged flight. It follows an earlier video showing Alef’s Model Zero Ultralight driving and flying, including a flight over a Tesla Cybertruck.
How do always-on AI agents actually work—and what makes them capable of handling tasks even when you’re not actively prompting them?
In a lecture for Stanford’s CS146S course, Lee Robinson, formerly of Vercel and Cursor and now part of SpaceX’s AI team, explains the engineering behind Grok Bot and the emerging architecture of persistent, proactive AI agents.
The 45-minute lecture covers:
→ What changed in models to make always-on agents possible.
→ The architecture behind persistent agents.
→ Harness engineering: the systems that enable models to use tools and take action.
→ Context engineering for long-running tasks.
→ Where agent architectures are headed next.
The central insight: An AI agent is more than a model. Its ability to work autonomously over extended periods depends on the surrounding infrastructure—tool use, memory, context management, and the harness that coordinates its actions.
A useful look at the engineering required to move beyond conversational AI toward agents that can work continuously, respond to events, and manage complex tasks over time.
Purpose Bevacizumab is an effective treatment for radiation necrosis; however, dose –dependent radiographic and edema response data are lacking. Methods Multicenter, retrospective cohort study of adult patients with primary central nervous system (CNS) tumors or brain metastases who received bevacizumab for radiation necrosis between June 2021 and July 2024. Patients were grouped by bevacizumab dose (7.5 mg/kg vs. 10 mg/kg administered every three weeks). Primary outcome was edema response, assessed by a blinded neuro-oncologist. Secondary outcomes included contrast enhancing disease response, clinical response, adverse events, treatment discontinuation, and re-dosing for recurrent radiation necrosis. Analyses were performed with and without adjustment for tumor type. Results 104 patients were included (7.5 mg/kg, n = 53; 10 mg/kg, n = 51). Tumor type differed significantly between cohorts (p < 0.001), with more primary CNS tumors in the 10 mg/kg group. Edema response differed between dosing groups in unadjusted analyses (p = 0.032), favoring the 7.5 mg/kg cohort; however, this association was attenuated after adjustment for tumor type. Contrast enhancing disease response did not differ between groups in either unadjusted or adjusted analyses (p = 0.43). Clinical response was higher in the 7.5 mg/kg group (65.2% vs. 37.8%, p = 0.013). Rates of adverse events, treatment discontinuation, and re-dosing were similar. Conclusions Taken together, both 7.5 mg/kg and 10 mg/kg dosing strategies achieved similar radiographic outcomes. Prospective, tumor-stratified studies incorporating edema-focused and patient-centered endpoints are needed to better define optimal dosing strategies in this heterogeneous population. These findings support an individualized approach to bevacizumab dosing in radiation necrosis and highlight the need for prospective, tumor-stratified studies.
NASA, space industry leaders, and the Johns Hopkins Applied Physics Lab discuss what it will take to construct a successful lunar outpost.
A: Small Language Models are typically built using three main approaches:
1. Training from scratch: The model learns language patterns from large, carefully curated datasets, with its parameters initialized and trained rather than inherited from an existing pre-trained model.
2. Fine-tuning a pre-trained model: An existing model is further trained on selected data to improve its performance on specific tasks, domains, or instructions.
3. Knowledge distillation: A smaller student model learns from the outputs, predictions, or other training signals of a more capable *teacher model*, transferring useful capabilities into a more compact system.
SLMs are often described as models with fewer than 10 billion parameters, although there is no universally agreed-upon size threshold.
These approaches can also be combined to build models that balance capability, computational efficiency, and deployment costs.
References:
