AI Roundup 126: Good Grok / Bad Grok

July 11, 2025.

Jul 11, 2025

Bad Grok

In another wild display of AI alignment (or lack thereof), early this week Grok began posting antisemitic messages, praising Hitler, and engaging in detailed descriptions of assault.

Between the lines:

A couple of nuances worth noting: the most widespread headline - Grok calling itself "MechaHitler" - appears to be the result of user prompting, and not a spontaneous decision from Grok.
However, there do appear to be significant differences in Grok's safety guardrails compared to other chatbots - simple prompt changes (system or otherwise) aren't going to get Claude to make antisemitic comments, for example.
Musk had previously complained that Grok was too "woke" and suggested in June that it would be "retrained." It's unclear whether this is a result of recent changes or simply the by-product of training an AI on X's content.
Ultimately, biggest casualty appears to be Linda Yaccarino, X's CEO, who resigned in the wake of the latest controversy.

Good Grok

Strangely enough, Musk's AI lab also released Grok 4 this week, a benchmark beating AI model available alongside a new $300/month subscription.

The latest stats:

Grok 4 is the company's first full-scale reasoning model, with a sizable 256K token context window and competitive API pricing. It comes in two sizes, "Grok 4" and "Grok 4 Heavy" - the latter isn't yet publicly available, but will be available first to those on the "SuperGrok Heavy" subscription plan.
Impressively, it scores better on several leading benchmarks, including newer ones designed to stump AI like ARC-AGI (66%) and Humanity's Last Exam (24%) - but as of yet there is no system card to get more details on its training and evaluation.
It also seems like more models are coming soon—the company said Wednesday that an AI coding model will be released in August, a multi-modal agent in September, and a video-generation model in October.
However, it seems that no version of Grok is without some drama - Grok 4 appears to consult Elon Musk's personal opinions and posts when answering controversial questions.

Elsewhere in AI anxiety:

The Internet Watch Foundation identified 1,286 AI-generated videos of child sexual abuse in H1 2025 globally, up from just two in H1 2024.
Hertz and other car rental agencies are increasingly relying on AI scanners that use high-res imaging to flag even tiny blemishes, angering some customers.
An impostor used AI to mimic Marco Rubio's voice and writing style to contact foreign ministers, a US governor, and a member of Congress.
Velvet Sundown, an AI-created band with 850K+ monthly Spotify listeners and two albums in June alone, shows how some listeners are turning to AI music.
And Missouri AG Andrew Bailey is investigating leading AI companies, claiming the companies' AI chatbots are discriminating against President Trump.

Apple picking

Another week, another batch of headlines on Meta's relentless recruitment. This time, it poached Ruoming Pang, Apple's head of foundation models.

The big picture:

Pang was reportedly offered a staggering (albeit unconfirmed) $200M+ over four years.
It's not just OpenAI and Apple - Waymo is yet another company targeted - but reporting suggests that there have been fewer defections from DeepMind and Anthropic than OpenAI.
What ultimately remains to be seen is whether Zuckerberg can get all of these incredibly accomplished folks to work well together: on the one hand, some are suggesting he's creating a "halo effect" that's luring other leading scientists to Meta; on the other, an outgoing researcher has called out a "culture of fear" at the organization.

Elsewhere in the FAANG free-for-all:

AWS is launching a marketplace for AI agents, and Anthropic will be one of its partners at the AWS Summit in NYC next week.
Google added image-to-video generation to Veo 3 in the Gemini app for Pro and Ultra subscribers, and users have created 40M+ videos since its launch.
Condé Nast and Hearst signed multiyear agreements with Amazon to license their content for use in Amazon's AI shopping assistant, Rufus.
Google launches AI ad tools in India following the repeal of the "Google Tax" which has made the market more attractive to global tech companies.
And Amazon is considering another multibillion-dollar investment in Anthropic as it seeks to position itself ahead of Google, which has also invested $3B+.

Comet

Perplexity launched Comet, an AI-powered web browser featuring integrated AI search and an assistant agent, as the startup aims to challenge Google's dominance.

Why it matters:

Perplexity is betting that controlling the browser is key to "infinite retention" - if users default to their platform, they can capture search queries without relying on Google Chrome's distribution.
As we've discussed, AI is reopening the Browser Wars: The Browser Company launched Dia, OpenAI is reportedly about to launch its own browser, and Google has been rapidly integrating AI features into Chrome.
While Comet's assistant shows promise for simple tasks like summarizing emails and answering questions about web pages, it still suffers from hallucinations on complex requests - highlighting that AI agents remain more novelty than reliable tool for demanding workflows.

Elsewhere in OpenAI:

OpenAI officially closed its ~$6.5B acquisition of Jony Ive's startup, now called io Products after a trademark dispute with hearing device startup Iyo.
The company reportedly overhauled its security, using biometric checks and isolating sensitive information to protect IP, such as model weights, from espionage.
OpenAI plans to release its first open-weight model since GPT-2 in 2019 - it may launch on Hugging Face and other platforms as soon as next week.

Thanks for reading Artificial Ignorance! This post is public so feel free to share it.

Things happen

Hugging Face is taking orders for its Reachy Mini robots at $449 and $299. FlexOlmo lets data owners remove training data. Clorox uses AI for ad creation in its $580M digital overhaul. Superintelligence is trendy among AI leaders despite being possibly overhyped. Moonvalley debuts Marey, its licensed AI video model. AI startups received 53% of global VC dollars in H1 2025. Fighting AI scraping with cryptographic challenges and AI Labyrinth. AI agents' sky-high ROI use case: Stealing crypto. Why I don't ride the AI Hype Train. On-the-job learning faces threats from hybrid work and AI. Hollywood video game actors get AI protections in new contract. The EU won't pause its AI Act rollout despite company objections. AI agent Manus relocated to Singapore. California's SB 53 gets amendments for AI transparency. India's AI independence scramble amid linguistic diversity challenges. LLM-assisted writing in biomedical publications. UK's Turing AI institute pivots to national security. Sam Altman is politically homeless. The right way to embed an LLM in group chat. Trump administration plans AI chip curbs on Malaysia and Thailand. Academic papers had hidden AI prompts for positive reviews. The EU publishes voluntary AI code of practice. Indeed and Glassdoor plan to cut 1,300 US jobs. Nvidia plans Blackwell RTX Pro 6000 for China. Microsoft's AI saved $500M+ in 2024. YouTube cracks down on AI slop. Microsoft's $4B AI education commitment. Google adds AI Mode to Circle to Search. Nvidia hits $4T market cap. Meta buys 3% of EssilorLuxottica. IBM unveils Power11. Huawei refutes code copying accusations. Gemini's personalized content shows Google's data leverage. Microsoft Azure AI services: $11.5B revenue. Google hires ad tech staff for publishers. AI impact on developer productivity. The force-feeding of AI on an unwilling public. MCP-B: AI Browser Automation Protocol. UK publishers file antitrust complaint against Google. OpenAI's stock compensation jumps 5x to $4.4B. Teachers' union opens AI training hub with $23M. DeepSeek's impact 150 days later.