SnacksCast
The SnacksCast from PodSnacks.com distills the most important ideas, debates, and insights from the top 200 podcasts into a sharp, under-15-minute listen. We analyze hours of conversations so you can stay informed on what the world’s smartest voices are saying—in minutes, not hours.
SnacksCast
AI & Tech: Cursor Composer 2.5 nears Opus on coding benchmarks at fraction of the cost
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Podsnacks! The news you need from the voices you trust. This is PodSnacks. Here's what's driving the conversation. Wednesday, May 20th, 2026. Cursor just released its in-house model called Composer 2.5. It scores within one point of Opus 4.7 and GPT 5.5 on tough coding benchmarks, yet charges only 50 cents per million input tokens and $2.50 per million output tokens. That is roughly 10 times better token efficiency than current leaders. The AI Daily Brief covered the release in detail. Anthropic's Mythos preview model takes a different approach. According to Cloudflare's review, it can combine multiple attack primitives into working exploits, generate proofs of concept, and refine failures iteratively. The behavior feels closer to a senior security researcher than a basic scanner. Google built a multimodal agentic scanning system called MDash. It deployed over 100 specialized agents across frontier models and found 16 vulnerabilities, including four critical remote code execution flaws in the Windows kernel, TCP, IP stack, and IKEV2 service. Those issues were fixed in Maze Patch Tuesday. SecurityNow reported on the findings. OpenAI introduced Daybreak. It's Frontier AI built specifically for cyber defenders. The system offers controlled access models ranging from standard GPT 5.5 up to fully permissive versions for red teaming and vulnerability discovery. Programmers are now overseeing 20 simultaneous AI coding bots that write and debug code. Mark Andreessen noted on the A16Z show that this setup produces 5 to 20 times more output per hour and turns top engineers into 24-hour AI vampires who rarely sleep. Life Rays agentic AI support system reads incoming tickets, breaks them into workflows using documentation and known issues, and delivers proper resolutions in 51% of complex developer cases while keeping humans in the loop. Blackstone committed $5 billion to a new AI cloud venture that will deploy Google's specialized chips and target 500 megawatts of capacity next year. Hive Digital announced a $2.5 billion Canadian AI gigafactory near Toronto with 320 megawatts capacity and 100,000 GPUs. Initial operations are already allocating 32 MW and expected online in 2027. Leopold Ashenbrenner's Situational Awareness Fund disclosed over $2 billion in semiconductor-related positions while trimming some core weave calls. The BET assumes sustained half an order of magnitude annual compute growth. Near-term Watt constraints should begin easing in 2027 and 2028 as new turbines and energy sources come online. Orbital compute racks and sun-synchronous orbits could solve the problem long-term through solar wings and laser-linked virtual data centers. The Democratic Republic of Congo and Uganda have recorded over 130 deaths and 500 suspected cases in this strain. Every study shows that covering GLP 1 drugs such as Ozempic does not reduce overall health spending for employers, even though one in eight U.S. adults now takes the medications. Connecticut spent $80 million last year on GLP 1 coverage for state employees while requiring food tracking, weigh-ins, and labs to limit access to committed patients. Every human wakes between 1 and 3 a.m. as core body temperature drops to prevent hypothermia. This causes a brief awakening that most people roll through and return to sleep without noticing. The 4-7-8 breathing technique inhale 4 seconds, hold 7, exhale 8 for 20 cycles, lowers heart rate below 60 by activating the parasympathetic response. Consumer sleep trackers are consistently inaccurate because they guess from proxies rather than brainwaves. Only week-long trends or deltas should be considered, never single-night scores that create self-fulfilling prophecies. Mouth taping is dangerous and unnecessary. Fix nasal congestion with saline, flonase, or internal dilators instead, since 90% of mouths close naturally during sleep. That's the rundown.