Here’s what’s on our plate today:

• 🧪 GPT-5 leans on test-time compute as users compare it to 4o.

• 🛡️ Companies draft AI agents into cyber defense.

• 🪦 ‘Deathbots’ stir an ethics fight over digital grief.

• 🔁 Altman addresses rollout and restores 4o after backlash.

Let’s dive in. No floaties needed…

The Laboratory

The GPT-5 gamble: Smarter routing, skeptical users

From bestselling novels to news outlets, artificial intelligence is everywhere. Conversations focus on AI’s impact, how it works, and how it can be used to boost efficiency. Against this backdrop, expectations of a new LLM, especially from OpenAI, the company that started the AI race, were bound to be high. To add to the hype, OpenAI CEO Sam Altman said the new version, although it cannot evolve continuously, could provide PhD-level expertise. That was enough to get users hooked. They expected the new to increase gains when used for generating code, writing, research, analysis, and problem-solving.

For OpenAI, the prospect of a more advanced model would assure investors that it could continue scaling AI models, pull in more paying subscribers, and boost investments by justifying the enormous sums of money it is spending to fuel development. During the release, OpenAI showed how GPT-5 could be used to create entire working pieces of software based on written text prompts, commonly known as "vibe coding." With a lot hanging in its success, after multiple delays, speculations, and anticipation, the GPT-5 model was released to all 700 million ChatGPT users.

The initial response to GPT-5

Within twenty-four hours of the launch, users had taken to social media to express their disappointment. Users on Reddit said they were unimpressed with the next generation of ChatGPT that used GPT-5, and demanded that OpenAI bring back older models, including the previous 4o and 4.1 models.

According to a report from TechRadar, a Reddit thread titled "GPT-5 is horrible" had nearly 6,000 upvotes and over 2,000 comments at the time of writing this article. The thread was full of users dissatisfied with the new release. Many users said they felt that the latest AI model was being used to force non-paying users to pay for a subscription. Users also complained that the new model had less personality and demanded that OpenAI make older models available, a request the company has given in to, albeit only for paid subscribers.

In a Reddit AMA, Altman relented, saying he would bring back the older 4o model, but would watch usage and determine how long to support it. Notably, ChatGPT Pro, Team, and Enterprise users already have an option to get back to older models by turning on the show legacy models banner in settings.

GPT-5 is an important step for OpenAI

OpenAI started the AI race when it first introduced ChatGPT to the world in 2022 as a research preview. At the time, users were mesmerized by the chatbot’s ability to generate coherent, context-aware answers, though some experts cautioned about its factual accuracy and overconfidence. The app amassed over a million users in just five days, and over 100 million users within two months of its release, becoming one of the fastest-growing apps ever.

The initial success of ChatGPT paved the way for the release of GPT-4, a large language model that made huge leaps forward in intelligence. By March 2023, GPT-4 launched, offering better reasoning, broader general knowledge, and improved safety guardrails. Industry analysts praised its jump in capability, especially for professional and academic tasks, though concerns about hallucinations persisted. An earlier version, GPT 3.5, received a bar exam score in the bottom 10%; GPT-4 passed the simulated bar exam in the top 10%.

OpenAI followed up on the release with GPT-4o in May 2024. The launch of successive models by OpenAI was driven by a mix of user demand, enterprise interest, and competitive pressure from rivals like Anthropic’s Claude, Google DeepMind’s Gemini, Meta’s LLaMA, and xAI’s Grok. However, while the AI startup was launching more advanced models, it was also struggling with the problem of scaling its growth in model size and capabilities.

The scaling squeeze

The success of AI models from OpenAI raised expectations for successive models. However, OpenAI was facing multiple challenges in scaling up. The first challenge was to find the compute needed for next-gen models. As rivals started releasing more models, the demand for high-end GPUs capable of running AI models skyrocketed.

A study from Cornell University found that training costs for frontier AI models (like GPT-4) have been increasing at a rate of 2.4× per year, projecting that by 2027, only the most well-funded players will afford such models. OpenAI tried overcoming the challenge by diversifying its hardware sources, renting Google’s AI chips to power ChatGPT and other products. Google’s Tensor Processing Units (TPUs) are cheaper compared to Nvidia GPUs, which the company had been using so far.

Apart from computing, Large Language Models also need training data, which OpenAI was finding increasingly difficult to source. According to a report from Reuters, OpenAI's former chief scientist, Ilya Sutskever, said that while processing power was growing, the amount of data was not. A research paper projects that the supply of public, human-generated text may be fully consumed by 2026 to 2032, after which model scaling using fresh, real data becomes increasingly difficult.

Though companies have the option of using synthetic data, it can lead to models degrading over time due to a lack of diversity, impacting performance, coherence. Models trained on synthetic data also run the risk of collapsing as AI would learn from its flawed outputs, amplifying errors and becoming less capable in edge cases.

OpenAI’s bet: Test-time compute

To reduce stress on its GPUs and show improvement in performance in its model without having to spend exorbitant amounts of chips, OpenAI is reported to have used the test-time compute method.

Older models like GPT-4 would spend a limited amount of compute to produce an answer, regardless of difficulty. However, under the test-time compute method, GPT-5 acts like a router. When a user inputs a prompt, the model first evaluates the difficulty of the problem, decides which model to use based on conversation type, complexity, tool needs, and explicit intent. And then, allows the model to run more thinking cycles before producing an answer. This allows GPT-5’s real-time router to automatically select the best model for each query, simplifying the user experience without manual choices, while saving compute on simpler queries.

This is the first time OpenAI is exposing this capability to the public. Earlier, such adaptive compute routing was only used internally for research or fine-tuning. The method is something Altman thinks is important to the company’s mission to build more powerful AI models, the investments in which he believes are still inadequate.

The stakes

In January 2025, U.S. President Donald Trump announced private sector investments of up to $500 billion to fund infrastructure for artificial intelligence. The investment adds to the billions already spent on developing and deploying AI models.  According to a Reuters report, returns on this investment are also rising with demand in internet search, digital advertising, and cloud computing powering revenue growth for giants like Microsoft, Meta, Amazon, and Alphabet. Big Tech is expected to further increase spending to meet soaring AI services.

In this scenario, OpenAI faces an uphill task. Whether GPT-5 cements its lead or marks the moment its rivals catch up will depend not on hype, but on how well it can balance ambition with scalability.

Quick Bits, No Fluff

• Apple’s next event: ‘iPhone 17’ rumored as the thinnest yet, with camera tweaks, lighter frames, and updates across other products.

• U.S. signals a softer stance on AI-chip sales to China; Nvidia and AMD get breathing room—but export scrutiny, geofencing, and politics still loom.

• OpenAI restores GPT-4o after user backlash; ChatGPT regains its fast, multimodal default while model shuffles continue.

Brain Snack (for builders)

Meme Of The Day

Rate This Edition