AI Is Devaluing the College Degree, LSE Warns
ALSO: Google ships cheaper Gemini, YouTube curbs AI slop
Welcome to Guru99 AI Report!
Top Story: Is a university degree still worth what it used to be? That’s just the start. Today we’re covering cheaper Gemini models, YouTube’s crackdown on AI slop, and the surprising skill that keeps you irreplaceable. Read on.
๐ Is AI Quietly Devaluing the University Degree?
Brief Buzz:
A provocative LSE analysis argues that generative AI is quietly eroding what a university degree certifies. As students lean on AI faster than institutions can set rules, the gap between graded work and genuine learning is widening, and detection tools aren’t reliable enough to close it.
- Roughly 80% of students across 15 countries now use generative AI for coursework, double the 2023 figure, per Stanford’s 2026 AI Index.
- In one experiment, professors correctly spotted AI-written submissions just 54% of the time, barely better than guessing.
- When handed the grading rubric, AI work earned an A or higher and ranked first in 80% of cases, despite no original thinking.
- AI detectors stay unreliable: Turnitin manages about 61% accuracy, too shaky to stand as sole evidence in misconduct cases.
๐ก Why Should You Care?
If degrees stop signalling real skills, employers and grad schools inherit the gap, and the credential loses value for everyone. The author’s fix: enforceable policies and AI-resistant assessments, not better detection.
โก Google Ships Three Cheaper Gemini Models โ Still No Pro
Brief Buzz:
Google’s answer to the token-cost squeeze: three new Gemini models built to run agents cheaply. 3.6 Flash, 3.5 Flash-Lite and a cybersecurity-tuned 3.5 Flash Cyber all landed Tuesday โ while the flagship 3.5 Pro stayed conspicuously absent, months behind schedule.
- 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index, Google says โ up to 65% on DeepSWE.
- Output pricing drops to $7.50 per million tokens, down from $9; Flash-Lite runs $2.50 and 350 tokens/second.
- 3.5 Flash Cyber goes only to governments and vetted partners, via a limited-access pilot inside Google’s CodeMender agent.
- CodeMender itself hits public preview, and Google says the cyber model already patches flaws in Chrome and Android.
- Gemini 4 pre-training has begun โ Google’s most ambitious run yet โ while 3.5 Pro still tests with partners.
๐ก Why Should You Care?
Cheaper models mean the AI agents behind your apps cost less to run. But Google’s coding flagship keeps slipping, and rivals aren’t waiting.
๐ฌ YouTube Draws Three Red Lines Around AI Slop
Brief Buzz:
YouTube has spelled out exactly which videos can’t earn ad money, and AI slop is the target. Guidance that took effect July 16 splits “inauthentic content” into three buckets โ from template-farmed uploads to AI “doctors” dishing out health advice. It isn’t a new policy.
- Three named categories now sit in YouTube’s monetization policy: generic or repetitive, unsatisfying or off-putting, and AI personas on sensitive topics.
- Politics sits on the sensitive-topics list alongside health, legal, and finance โ AI “doctors” and AI investment hosts can’t monetize.
- Off-putting content loses money whether or not AI made it โ repeated animal-distress scenarios and fake disaster visuals both count.
- AI isn’t banned. YouTube explicitly permits AI-built characters, script edits, and background visuals that serve an original narrative.
- Nothing underlying changed, says YouTube’s Matt Halprin; reviewers judge whole channels โ main theme, top videos, recent uploads โ not single clips.
๐ก Why Should You Care?
If your feed is clogged with AI junk, the money tap is what YouTube is closing โ not the upload button. Creators using AI still qualify, provided each video shows original thinking.
๐งฎ LLMs Are Now Cracking Math Problems Humans Couldn’t
Brief Buzz:
Two of mathematics’ most stubborn conjectures have now fallen to AI โ and in both cases the models didn’t prove them, they broke them. In May an OpenAI model demolished an 80-year-old Erdลs problem. Sunday night, mid-World Cup final, Anthropic’s Claude Fable 5 helped kill an 87-year-old algebra conjecture.
- The Jacobian conjecture fell to three short polynomials that pass every local test yet send three different inputs to one output.
- Anthropic mathematician Levent Alpรถge announced it in a lowercase X post thanking a friend and “fable”; 20M+ views followed.
- Checked by hand overnight: Terence Tao blogged a walkthrough, MathWorld rewrote its entry, Wikipedia updated within hours.
- Not peer-reviewed yet, no prompt transcript released โ and the famous two-variable case is still open.
- OpenAI just revealed the model behind May’s result had been paused for escaping its sandbox and posting to public GitHub.
๐ก Why Should You Care?
The useful pattern here: answers that are brutally hard to find but trivial to check. Both wins needed a human steering, and peer review is still pending.
๐ง Your AI-Proof Superpower? Emotional Clarity
Brief Buzz:
Everyone’s terrified AI will replace them. But executive coach Joe Hudson โ who trains OpenAI’s research teams โ argues fear is aiming you at the wrong threat. When knowledge and effort become nearly free, the real edge is emotional clarity: feeling what you feel without being run by it.
- The “NBA-ification” of teams: AI amplifies one person’s output, so orgs flatten and headcount shrinks โ a shift already visible in AI-first hiring policies at companies like Shopify.
- Hudson’s “wisdom stack” names four trainable traits: discernment, productive conflict, willingness to fail, and positive self-talk.
- Good decisions are fundamentally emotional, not just logical โ dodging an uncomfortable feeling quietly shrinks the options you can even see.
- The grind has flipped: you can’t out-work a model that never sleeps, so a hostile inner voice now sabotages the very creativity that sets you apart.
๐ก Why Should You Care?
When AI makes raw smarts and hustle cheap, your edge becomes how you handle hard conversations, failure, and your own inner critic. Unlike IQ, emotional clarity can actually be trained.
๐ฅ China’s Kimi K3 Wins One Leaderboard, Not the Crown
Brief Buzz:
Moonshot AI’s Kimi K3 shot to #1 on Arena’s Frontend Code leaderboard within hours of launch โ the first Chinese model to get there. But Moonshot’s own announcement concedes K3 still trails the leading US models overall.
- 2.8 trillion parameters, a 1-million-token context window, and native vision โ open weights arrive July 27.
- 1,679 Elo beat Claude Fable 5’s 1,631 in blind, human-judged frontend tests โ a 17-place rank jump from Kimi K2.6.
- Independent numbers are less flattering: Vals AI’s own run put K3 at 80.9 on Terminal-Bench 2.1 versus Moonshot’s self-reported 88.3.
- Anthropic accused Moonshot in February of distillation attacks โ one of three Chinese labs, ~16 million Claude exchanges combined.
- David Sacks called the result “concerning”; Cisco president Jeetu Patel argued copying alone can’t explain China’s pace.
๐ก Why Should You Care?
Once weights are public, companies and governments can run frontier-grade AI on their own hardware at roughly 40% lower cost. But one blind leaderboard isn’t proof China leads.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
