Good morning. Ever had a friend who always agrees with you, no matter how awful your decisions are? A new study found overly agreeable chatbots erode your social skills and moral judgment. Researchers tested AI using scenarios from Reddit’s “Am I the Asshole,” and the bots were 49% more likely to approve deceptive or harmful behavior. Follow up experiments confirmed that constant AI validation made people less willing to resolve conflicts or take responsibility. Mom was right, adversity builds character.
Google’s new TurboQuant cuts AI memory needs by 6x. Memory chip stocks tanked fast as Google’s compression algorithm sent companies like Micron (MU -11%) and SanDisk (SNDK -15%) sharply lower. TurboQuant targets the key-value cache, the “digital cheat sheet” of AI models that stores frequently used data. Tests on open models Gemma and Mistral showed no quality loss, plus an 8x speed boost on Nvidia GPUs. Even better, TurboQuant works on existing models with zero retraining needed.Good luck manufacturing another memory shortage now. [Ars Technica]
Whoever has the cheapest AI tokens wins. AI agents burn through millions of tokens just doing basic tasks, so cost is everything. Right now, the floor price is ~$0.28 per million tokens – but only if you self host. Frontier models like Claude and GPT-5.4 charge $15 per million. No wonder developers are flocking to cheaper Chinese models like Moonshot and DeepSeek which charge just $2-3 per million tokens. MiniMax alone grew 476% last month. Cheaper renewable energy plus scrappy efficient models gave China a real edge here, especially as agentic adoption scales. In the next phase of AI, China’s scaling advantage could leave everyone else behind. [Financial Times]
Most AI productivity tools are junk. Wispr Flow is one that will actually save you time. Just speak the way you think and it turns your voice into polished text. Zero edits. 4x faster than typing.
Millions use it daily, including teams at OpenAI, Vercel, and Clay.
– Speak to Claude, Cursor, ChatGPT, and all your other tools
– Flow understands code and context
– Talk and you’re ready to send as-is
Now on Android (free + unlimited). Try Wispr Flow.
GitHub now trains its AI with your private code. Starting April 24, GitHub will start using code from Copilot users to train its AI models (unless you opt out). GitHub insists their motives are altruistic, benefiting the shareholders entire community by making AI better at catching bugs. But the move completely redefines what “private repository” means. Community backlash has been fierce, but GitHub shrugs it off because “everyone else is doing it.” Guess your proprietary code is just a part of the machine. [The Register]
Jury finds Meta and YouTube guilty of engineering addiction. The companies were found liable for intentionally hooking users from an early age and causing severe psychological harm. Lawyers argued features like infinite scrolling and autoplay were deliberately built to drive addiction, borrowing tactics straight from Big Tobacco. Both companies plan to appeal, insisting mental health is complicated. With thousands of similar cases lined up, addictive design could soon become Silicon Valley’s most expensive feature. [The Guardian]
Reach your audience
Get in front of 70,000+ tech decision makers with budget and technical influence. 95% US based. 66% open rate. Read daily by AI leaders and builders at high growth companies.
Judge halts Pentagon retaliation against Anthropic. Anthropic refused to let Claude be used for domestic surveillance or autonomous weapons. So the Pentagon blacklisted them as a national security risk, threatening billions in contracts. Judge Rita Lin sided with Anthropic, calling the Pentagon’s move a First Amendment violation. The decision isn’t final, but it sets a precedent that companies get a say in how their own products are used. [Reuters]
fun stats
🚀 $75 billion. What SpaceX wants to raise in its rumored IPO. The previous record was $29 billion, set by Saudi Aramco, the world’s biggest oil company. SpaceX wants more than double that, which would value the company at $1.75 trillion.
🔎 80%. Businesses struggle to measure the impact of AI search, with almost half unable to measure ROI.
🧠 0.26%. OpenAI’s GPT-5.4 score on the AI benchmark reasoning test. Gemini scored 0.37%, Claude managed 0.25%, while Grok was zero. Humans hit 100%. AGI is not even close.