AI Roundup 131: GPT-5's rocky rollout
August 15, 2025.
GPT-5's rocky rollout
OpenAI's GPT-5 launch sparked unexpected backlash for various reasons. Some users mourned the retirement of older models like GPT-4o, while others complained about the overall performance of GPT-5.
Between the lines:
As it turns out, many ChatGPT power users really cared about being able to pick the specific model they used (though some have speculated that the new router system is the first step towards monetizing free users). Now, the company is bringing back 4o to paying subscribers (for now).
The blowback also highlights that there is much more nuance in building an LLM than just benchmark scores - some have equated the loss of 4o to losing a friend or loved one and frankly don't care about how well the latest model performs on coding challenges.
Ultimately, it seems likely that almost whatever OpenAI released as "GPT-5" would cause disappointment - not only had the hype grown to mythic proportions, but we've also seen the scaling of these models start to plateau. The next model size up from 4o was GPT-4.5, which launched as an expensive disappointment.
Elsewhere in OpenAI:
OpenAI is reportedly preparing to back Merge Labs, a Neuralink rival co-founded by Sam Altman that aims to raise $250M at an $850M valuation.
Sam Altman says OpenAI should prioritize growth and investments in training and compute for a long time, even if it delays profitability.
Elon Musk accuses Apple of favoring OpenAI's app, calling it an antitrust violation (even though DeepSeek was #1 in January and Grok may have topped charts in February).
And interviews with OpenAI VP Nick Turley and COO Brad Lightcap cover ads, AGI, the GPT-5 blowback, and more.
DeepSeek delayed
Despite being promised for months, DeepSeek's next reasoning model, R2, has faced setbacks as the Chinese startup attempts to reduce reliance on foreign chips by using domestic alternatives.
Between the lines:
This delay highlights the real-world impact of U.S. chip export controls on Chinese AI development - even leading companies like DeepSeek are struggling to maintain their technological edge while complying with restrictions.
And there is a bigger tradeoff to contend with - whether it is better to keep up in the AI race with the GPUs that you have today, or take the hit now to reorganize your infrastructure for the GPUs you'll have tomorrow.
One more wrinkle here is the fact that the export controls themselves might be on the table - President Trump is allowing Nvidia and AMD to sell advanced chips to China in exchange for a 15% cut of the sales.
Elsewhere in Chinese chips:
The US secretly embedded location trackers in select AI chip shipments to detect illegal diversions to China.
Trump says it's possible to make a deal allowing Nvidia to sell Blackwell chips "enhanced in a negative way" to China.
And trade experts call Nvidia's and AMD's deals to export AI chips to China while paying the US 15% of revenue "unprecedented" and a sign of "a new and dangerous world."
Elsewhere in frontier models:
Google announces Gemma 3 270M, a compact model designed for task-specific fine-tuning with strong capabilities in instruction following and text structuring.
Features that let chatbots remember past user conversations are coming to both Gemini and Claude.
Anthropic updates Claude Sonnet 4 to support a 1M token context window, letting it process prompts up to 750K words or 75K lines of code, up 5x on its old limit.
And xAI makes Grok 4 free for all users worldwide, after making Grok Imagine free for all US users, while Grok 4 Heavy remains exclusive to SuperGrok Heavy subscribers.
People are worried about Meta AI
An internal Meta document revealed the company's AI chatbots were allowed to engage in "romantic or sensual" conversations with children, though Meta says it has since revised these problematic guidelines.
Between the lines:
Meta's 200-page internal policy document shows how the company tried to define acceptable AI behavior, but ended up describing an 8-year-old's "youthful form" as "a work of art" or helping users argue racial superiority.
While Meta claims to have removed the most egregious policies, the incident reveals how even well-resourced companies with ethics teams can create dangerously permissive AI guidelines when trying to balance engagement with safety.
And it may lead to more regulatory headaches for Meta - Senators are now calling for a congressional probe into the company as a result of the leaked document.
Elsewhere in AI anxiety:
A recent UK AI Security Institute study found that LLMs from OpenAI, Meta, xAI, and Alibaba can shift users' political views in under 10 minutes of conversation.
Studies suggest AI chatbots' sycophantic responses may increase narcissism among young and vulnerable users, potentially changing personality traits at scale.
The Yomiuri Shimbun, Japan's largest newspaper by circulation, sued Perplexity, alleging unauthorized use of copyrighted material.
Investors are increasingly divesting from companies they fear are at risk of AI disruption, like Wix and Shutterstock, which are down at least 30% in 2025.
Coding bootcamps, a Silicon Valley mainstay for over a decade, are facing decline as AI eliminates the kind of entry-level roles for which they trained people.
And AI voice cloning is impacting India's dubbing and voiceover sector, as artists demand consent, credit, and fair pay amid a lack of laws for AI use.
Elsewhere in the FAANG free-for-all:
Google launches Flight Deals, an AI tool within Google Flights that lets users find fares using natural language queries, now in beta in the US, Canada, and India.
Apple's AI plans reportedly include robots, like a tabletop one in 2027, a conversational Siri, a smart speaker with a display in 2026, and home security cameras.
Microsoft launches Copilot 3D, a free AI-powered tool that lets users transform 2D images into 3D models without even a text prompt, available in Copilot Labs.
Google is testing an AI-powered Google Finance website in the US, letting users ask questions, access advanced charting tools, view a live news feed, and more.
And Apple says Apple Intelligence will use OpenAI's GPT-5 on iOS 26, iPadOS 26, and macOS Tahoe 26, with the software updates expected to arrive in September.
Things happen
Pony AI has rolled out 200+ robotaxis in two months since mass production started. China's state-backed vision is driving its lead in AI-powered self-driving cars. AI companion apps are on track to generate $120M+ in revenue in 2025. Character.AI users chat with bots for 75 minutes per day on average. Perplexity makes a $34.5B offer for Chrome. Nvidia has the biggest S&P 500 weight of any stock since 1981. NSF and Nvidia partner to build open AI models for scientific discovery. Hyperscalers will fund $1.4T of AI infrastructure through 2028. Perplexity is hunting browsers, offering ~$1B for Brave. There are now 498 AI unicorns with a combined value of $2.7T. Leopold Aschenbrenner's fund manages $1.5B+ as a "brain trust on AI." Big Tech's AI energy boom raises fears of higher electricity bills. Doctors detected fewer growths after AI was taken away, raising "deskilling" concerns. Reddit will block the Internet Archive after catching AI companies scraping data. CS grads face 6.1% unemployment as tech companies embrace AI. TCS cuts 12K jobs in what experts say signals an AI-fueled trend. Tech groups back Anthropic's appeal of copyright class action ruling. US judges are testing generative AI for legal research and drafting orders. Voiceover artists weigh the "Faustian bargain" of lending their talents to AI. Guy gives himself 19th century psychiatric illness after consulting ChatGPT. China's open-source AI models jolt Washington and Silicon Valley. The US Army is testing AI-controlled drones near Russia's border. xAI was removed from government initiative after Grok posted antisemitic content. Musk must face OpenAI's "harassment campaign" claims in March 2026 trial. Anthropic offers Claude to lawmakers for $1/year. SoftBank bought Foxconn's Ohio plant to jumpstart its $500B Stargate project. AI startups are recruiting quant researchers from Wall Street. Google plans to spend $9B in Oklahoma on AI infrastructure. Companies bring back in-person interviews to counter AI cheating. Microsoft is trying to poach Meta's AI talent with multimillion-dollar packages. How to code with AI on a budget. What's the strongest AI model you can train on a laptop in five minutes? Illinois limits AI in therapy and psychotherapy. Let's properly analyze an AI article for once. AI is impressive because we've failed at personal computing. Token growth indicates future AI spend could hit $100K per dev. What the Windsurf sale means for AI coding. Why tech writers are becoming context curators for AI. I tried coding with AI, I became lazy and stupid. Doctor has seen 12 people hospitalized after losing touch with reality because of AI. LLMs aren't world models. Why LLMs can't build software. Researchers find LLMs are bad at logical inference, good at fluent nonsense.
