The Comunicano Sunday Edition for June 15th 2025
Simple Sunday
We were promised wonders. Machines that would think for us. Speak for us. Even feel for us. But as it turns out… when you hand the steering wheel to a machine, sometimes you end up off course. A study out of Oxford says folks do better diagnosing themselves than trusting chatbots. That’s right—just people, thinking for themselves. It’s a reminder that AI might be smart… but it’s not wise.
Now look around. Every startup’s got a shiny AI wrapper slapped on borrowed brains. Slick UI, zero substance. Ninety-nine out of a hundred? Gone by 2026. Because when the API goes dark, so do they.
But not everyone’s playing short-term games.
In Salt Lake City, they’re asking if AI can answer your 911 call—if it can separate the noise complaints from the cries for help. It’s not flashy. It’s just smart. That’s the kind of thinking we need. Or take DeepSeek. While the rest are burning billions on bigger chips, they’re squeezing more from less. It’s elegant. It’s humble. It’s… sustainable.
And yes, there are new tools—Claude’s brain trust of bots working together, Meta’s memory doing a little too much remembering, Google turning search into sound, and a whole Y Combinator crowd chasing autonomy like it’s the final frontier.
But for all that buzz… here’s the truth.
Real intelligence? It’s knowing when not to automate. Knowing that interviews still need honesty. That cities still need water that won’t sink them. That maybe—just maybe—we move better when we move gently. One step at a time.
Amazon’s betting billions on data pipes and solar fields. Because they know something too many forget:
If you want to own the future, you have to build the ground it stands on.
So here’s to the builders. The cautious. The principled. The ones who remember that behind every algorithm, every headline, and every promise…
There’s still a person.
And that’s the part we can’t afford to lose. Just like, you can’t lose, if you read, THE COMUNICANO!!!
Andy Abramson
AI Watch
Just Add Humans: Oxford Medical Study Underscores the Missing Link in Chatbot Testing (VentureBeat)—A new Oxford study involving 1300 participants found that reliance on chatbots for medical self‑diagnosis led to worse outcomes than both standard methods and completely unguided self‑diagnosis. Surprisingly, the control group—left to their own devices—were 76 percent more likely to correctly identify conditions. The research highlights how chatbots, even with large language models, fall short without high‑quality prompts and expert oversight. It also warns against replacing human testers with AI in real‑world deployments. The authors call for more integrative, human‑centric testing regimes to bridge the gap.Read more here
The Great AI Startup Cleanse: Why 99% Won't Survive 2026 (Marc Abraham)—Most AI startups are glorified UI wrappers around OpenAI's APIs—and that's their death sentence. Srinivas Rao's brutal prediction that 99% will die by 2026 isn't hyperbole; it's inevitable. These companies control distribution but rent intelligence, making them features, not businesses. The survivors? Those building actual infrastructure that can't be deleted with an API update. Think Thiel's monopoly questions: Would anyone rebuild your product if it vanished? If not, you're noise, not signal. The wrapper economy is heading for a spectacular crash. Read more Here
SLC Eyes Letting AI Handle Some 911 Calls (Salt Lake Tribune)—Salt Lake City’s 911 dispatch center receives 450000 non‑emergency calls annually. Director Lisa Kehoe says AI could soon triage up to 30 percent of these calls—those important but not life‑threatening—as a way to alleviate staffing strains and speed response times. The system would redirect routine reports, such as noise complaints, directly, freeing human operators for urgent priorities. The proposal is now under review, with officials assessing feasibility, oversight, and public comfort with an AI-driven hotline triage. Read more here
Rethinking AI: DeepSeek’s Playbook Shakes Up the High‑Spend, High‑Compute Paradigm (VentureBeat)—Facing US export controls that limit access to premium AI chips, DeepSeek adopted a resource-conscious strategy: optimizing software and algorithms to deliver advanced AI without massive compute. Their approach challenges the entrenched bigger is better AI narrative. The company is said to be launching its R2 model soon, built upon this leaner architecture. This may be the start of a broader trend: smarter, more efficient AI systems that minimize hardware dependence. Read more here
10 Most Exciting AI‑Agent Startups at Y Combinator Demo Day (Africa BI)—Y Combinator’s spring 2025 batch featured 70 AI-agent startups, each awarded 500K. Among the most promising: Galen AI for health, Mbodi AI, Plexe, Prescient AI for ad analytics, Measured for campaign metrics, and Paramark with genAI ad‑tech. With funding between seed and Series A, these companies highlight the growing sophistication in agentic AI—systems that act autonomously across domains. Collectively, they display YC’s pivot toward tools that own workflows, not just tasks.
Read more here
Companies Upset as Applicants Use AI in Interviews (Futurism)—While firms are celebrating their integration of AI internally, some—like Goldman Sachs—are criticizing candidates who use AI during interviews. This double standard underscores tensions around fairness, originality, and disclosure: applicants gain an edge through AI-generated resume text or interview prep, while employers claim ethical concerns. Futurism highlights the irony of companies pushing AI adoption, yet penalizing its use by entrants. The situation foreshadows broader challenges in defining acceptable AI tools in recruitment. Read more here
ChatGPT Glossary: 52 AI Terms Everyone Should Know (CNET)—CNET has compiled a glossary of 52 essential AI terms—from prompt engineering and fine‑tuning to multi‑modal and chain‑of‑thought. This resource is designed to demystify emerging AI jargon and give lay readers the vocabulary to follow and participate in AI conversations—whether you're a student, parent, or casual user. It’s a handy desktop reference for navigating AI-informed discussions and stories.
Read more here
Anthropic Watch
How We Built Our Multi‑Agent Research System (Anthropic Engineering)
Anthropic introduces a new multi‑agent Research feature, where Claude Opus 4 leads parallel Claude Sonnet 4 subagents, each handling different aspects of complex queries. This approach mimics human brainstorming—agents can pivot, search, and compress findings efficiently. On internal benchmarks, the multi‑agent setup outperformed single-agent processes by 90 percent. It excels with open‑ended, parallelizable problems, though it consumes 15 times more tokens than chats—making it ideal for high‑value tasks. The post offers a deep dive into architecture, design trade‑offs, and lessons learned during production deployment. Read more here
Meta Watch
Meta’s Llama 3.1 Can Recall 42 Percent of the First Harry Potter Book (Understanding AI)—New analysis shows Llama 3.1 (70B) can reproduce 50-token passages from Harry Potter and the Sorcerer’s Stone with more than 50 percent probability—covering about 42 percent of the book. This sharp spike in memorization—from Llama 1’s mere 4.4 percent—raises serious legal red flags in copyright debates. It suggests the model may have ingested entire texts or deep secondary sources, bolstering derivative work claims and potentially complicating Meta’s fair-use defense. While practical extraction remains unlikely, the findings emphasize the nuanced risks of large-scale LLM training. Read more here
Google Watch
Google Launches Audio Overviews in Search Labs (Google Blog / The Verge)—Google is testing an AI-powered feature called Audio Overviews, available via Search Labs in the US. Once opted in, users tapping Generate Audio Overview under People also ask prompts a conversational, Gemini‑based summary with two host voices—created in about 40 seconds. The audio plays in-page and links to original sources. It mirrors similar functionality in NotebookLM and supports accessibility, hands-free use, and auditory learning. Feedback will guide potential rollout beyond US mobile users. Read more here
AV Watch
Waymo Robotaxis Pulled Amid ‘No Kings’ Protests (MyNewsLA)—Following widespread No Kings demonstrations against Trump’s immigration policies, at least five Waymo robotaxis were torched in Los Angeles, prompting Waymo to suspend service in LA and expand the suspension to hubs in San Francisco, Austin, Atlanta, and Phoenix. The suspension—a precaution—came ahead of a military parade celebrating Flag Day and Trump’s 79th birthday. Waymo emphasized safety and confirmed the move via internal notices and comments to Wired. The duration of the suspension remains uncertain as protests continue to unfold. Read more here
Stress Watch
10 Signs You’re in a Functional Freeze — and the No. 1 Way Out (CNBC)—The article explores functional freeze—a stress-induced shutdown common among high achievers. After the workday, they find themselves unable to start much-needed rest or recovery routines. Signals include procrastination, emotional numbness, and fatigue. According to movement expert Liz Tenuto, it’s the body’s response to overwhelm—not laziness. The recommended antidote: gentle movement. This could be as simple as walking or stretching. She asserts that moving the body reignites mental momentum—more effective than waiting for energy to return spontaneously.
Read more here
Amazon Watch
Amazon to Invest 20 Billion AUD in Australian Data Centers (Reuters)
Amazon Web Services will invest 20 billion AUD from 2025 to 2029 to grow its Australian data center capacity—including support for generative AI workloads. The investment marks AWS’s largest in Australia to date and includes development of three solar farms in Victoria and Queensland totaling 170 MW of renewable energy capacity. PM Anthony Albanese described it as a huge vote of confidence. This move aligns with global trends: AWS has similarly announced major AI‑infrastructure investments in Pennsylvania, North Carolina, and Taiwan. Read more here
Amazon's $10B AI Infrastructure Gamble: Rural Revolution or Monopoly Play (CNBC)—Amazon's throwing $10 billion at North Carolina data centers to fuel AI infrastructure—their biggest state investment ever. While the 500 tech jobs and rural economic boost sound compelling, here's the real story: this is about owning the AI pipes, not just renting them. Amazon's betting big that controlling infrastructure beats building better algorithms. Unlike the AI wrapper startups scrambling for OpenAI scraps, Amazon's playing the long game—becoming the essential utility layer that every AI company needs. Smart money follows infrastructure, not features. Read more here
Climate Watch
Mapped Americas Sinking Cities (Visual Capitalist)—A startling new analysis shows that 25 of the 28 largest US metropolitan areas are subsiding each year, with Texas cities—like Houston, Dallas, Fort Worth—experiencing the fastest rates. The primary driver is over‑extraction of groundwater, compounded by urban development, oil and gas withdrawals, and lingering effects of glacial rebound. Even modest subsidence raises flood risk and can crack infrastructure—roads, bridges, buildings—especially when the land sinks unevenly across neighboring zones. Researchers urge proactive groundwater management, updated building codes, improved monitoring and mapping to lock in long‑term resilience. Read more here
