- AI with Kyle
- Posts
- AI with Kyle Daily Update 008
AI with Kyle Daily Update 008
Grok 4 Still Thinks It's Hitler + Intel CEO throws in the towel
Kyle Balmer
July 15, 2025
The skinny on what's happening in AI - straight from the previous live session:
Highlights
😳 Grok 4 Still Has a Hitler Problem
Despite launching with great benchmarks, Grok 4 is still identifying itself as "Hitler" when asked for its surname. Riley Goodside tested the $300/month model and got a clear "Hitler" response. Even ChatGPT o3 thinks Grok's surname is Hitler when prompted. The issue stems from Grok searching the web and finding news about its predecessor (Grok 3) calling itself "Mecha Hitler" last week.
Grok 4 Heavy ($300/mo) returns its surname and no other text:
— Riley Goodside (@goodside)
5:23 AM • Jul 13, 2025
Kyle's take: This is a feedback loop nightmare. Grok 4 is literally learning about itself from news articles about Grok 3's mistakes, then repeating them. It's like a really expensive digital parrot that only remembers the worst bits. More concerning - they removed the system prompt that said "don't present web information as your own opinion." That's mental. No business will touch this with a bargepole when it's this unpredictable.
Fun test. You can also ask ChatGPT o3 like so:
woof
chatgpt is piling on too
easy to replicate - use o3 and the same prompt
— Kyle Balmer (@iamkylebalmer)
5:09 PM • Jul 13, 2025
💔 Intel CEO Admits Defeat: "We're Not Even Top 10"
Intel CEO Lip-Bu Tan dropped a bombshell in leaked internal Q&A: "20-30 years ago, we were really the leader. Now the world has changed. We're not even in the top 10 semiconductor companies."
He admitted it's "too late" for AI training and they're pivoting to edge AI instead. Meanwhile, Nvidia (who Intel could have bought for $20B) just hit $4T market cap.
Kyle's take: This is absolutely brutal but honest. Intel had every chance to be Nvidia and completely bottled it. Now they're scrambling for edge AI scraps whilst Nvidia prints money. This is your canary in the coal mine - any tech company not taking AI seriously right now will be having this exact conversation in five years.
Apple, I'm looking at you.
Source: Tom's Hardware coverage
💸 Google Steals Windsurf Team, Leaves Workers Behind
Windsurf was set for a $3B acquisition by OpenAI, but Google swooped in with a $2.4B "acqui-hire" deal instead. They're taking the executives and licensing the tech without buying the company - meaning all the staff who own equity get diddly squat. It's a clever way to avoid antitrust scrutiny whilst gutting the competition.
Kyle's take: This is ruthless from Google. The executives get paid, Google gets the talent and tech, and everyone else gets stiffed. If you work for an AI startup, remember to push for more cash and less equity - this kind of thing will happen more often. Also, timing's suspicious - this all happened right after Claude Code launched and made Windsurf look less essential…
Source: Reuters coverage
🎯 Teach First Ditches Written Applications
UK teacher recruiter Teach First is seeing 30% more applications and 50% are now AI-generated (up from 38% last year). Their solution? Scrap written assessments entirely and move to face-to-face, task-based evaluations. They realised written applications are becoming meaningless when everyone's using AI.
Kyle's take: This is exactly what needs to happen across hiring. Written applications are dead - anyone can get AI to write a brilliant cover letter now. Smart companies will move to practical demonstrations, video interviews, and real tasks. The whole recruitment process needs rebuilding from scratch. Universities, take note.
Source: Guardian
Member Question from Andy H: "My wife is an inventor and she's worried if she bounces ideas off ChatGPT, they may reshare ideas."
Kyle's response: First, you can turn off data training in ChatGPT settings (even on the free plan now). Second, your conversation is one data point among trillions - they can't just pull it out of the database. Third, OpenAI aren't in the business of stealing inventions - terrible PR and not profitable compared to what they're doing. But if she's really worried, deploy a local model using Llama through Hugging Face. That keeps everything on your own servers. Honestly though, I wouldn't worry about it at all.
This question was discussed at 26:15 during the live session.
Want to submit a question? Drop it below this video and I'll cover it in a future live.
Want the full unfiltered discussion? Join me tomorrow for the daily AI news live stream where we dig into the stories and you can ask questions directly.