OpenAI Releases ChatGPT-5 and Open-Source Models

This TechCrunch article looks at the debate after it was discovered that AI firm Perplexity has been scraping content from Cloudflare websites without permission, and even having initially denied doing so. Cloudflare websites use the robots.txt file to block bots crawling their sites. This plaintext file specifies which parts of the website which bots (e.g., Googlebot, Bingbot) may have access to. In the case of Perplexity, the company’s bot was masquerading as a “generic browser intended to impersonate Google Chrome on macOS”. The robots.txt technique to block bots is clearly no longer appropriate. Further, even techniques like CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) are weakening because AI agents are able to solve some. Cloudflare is supporting the Web Bot Auth standard being developed by the Internet Engineering Task Force that will imposes a cryptographic-based identification on AI agent requests.

The issue of AI bots scraping is hotly debated. Gartner has predicted that search engine volume will drop by 25% by 2026, as users use chatbots to get referral to sites. The article cites a report that found that bot activity now outnumbers human activity on the Internet, and malicious bots now account for 37% of Internet traffic. Website owners will lose ad revenue from this progression because content is directly furnished by the AI chatbot. It remains to be seen whether websites will buy into the agentic AI paradigm (where agents purchase objects over the Web on a human’s behalf) for all of these reasons.

This article by Clément Delangue, co-founder and CEO of Hugging Face, calls on US Big Tech to embrace open-source language models. It follows the AI Action Plan by President Tump last week that included the priority to “encourage open-source and open-weight AI”. Delangue argues that success of DeepSeek-R1 is also due to it being released as open-source. The model became the most popular model on Hugging Face within days. The open-source and open-weight nature of Chinese models has led to a scientific boom around AI in China. In the US, earlier breakthrough models from Google, OpenAI and Stanford were open-source, but since 2020, all major models are proprietary. As a result, even US scientists are turning to Chinese models for their work on AI. Delangue writes “when openness slows down, the entire ecosystem follows”. Delangue also argues that open models are essential for transparency and audit, and to remove dependencies on commercial enteprises.

Remaining on the subject of open-source AI models, OpenAI is releasing two models under a permissive Apache 2.0 license. The model family is “gpt-oss” and is available in two sizes. The smallest is supposed to be able to run on a machine with 16 GB if RAM – equivalent to a powerful Macbook. The new models reportedly score similar results on benchmarks as the proprietary o3-mini and o4-mini models. The release of the models comes at a time when Chinese open-source models like DeepSeek’s R1, Kimi K2, and Alibaba’s Qwen family have become more popular than US open-source models. There is also the argument that Chinese models echo the political viewpoints of the Chinese government, suppressing facts such as the Tiananmen Square massacre in 1989. Western models have the possibility of carry other democratic values. The article also points out the OpenAI may be buying political leverage with the new US administration which, in its recent AI Action Plan, called for the development of open-source AI.

As part of its 1.4 trillion USD investment in AI before 2030, China is pushing for AI integrated healthcare. Tsinghua University launched the world’s first AI hospital in 2024 as a real-world pilot with 14 AI doctors. In November, it added “Zijing AI Doctor” which is a system with 42 AI doctors which cover 21 clinical specialties and over 300 diseases. In addition, it created 10’000 synthetic patients for testing, and the hospital management claims to be able to treat patients with 93% accuracy within a few days – something that would take real doctors years to complete. After internal testing, public testing began earlier this year and the Tsinghua Agent Hospital was opened to the public in April of this year. The current capacity of the hospital is 1’500 beds and there are 10’000 outpatients daily. Working in tandem with human doctors, the AI doctors offer real-time advice and streamline workflows on topics that include ophthalmology, respiratory medicine and radiological diagnostics. DeepSeek AI is the main model provider for Chinese hospitals today, being used in more than 260 hospitals. At Ruijin Hospital, DeepSeek models process 3’000 pathological slides daily. That said, AI is only used today in 0.7% of Chinese healthcare systems, mainly in urban areas. Officials are afraid of a healthcare divide in quality between urban and rural regions.

OpenAI has released ChatGPT-5 which the company calls its first “unified” model, combining reasoning capabilities from the o-series models and fast response times of the GPT series models. OpenAI claims that the model performs very well on standard benchmarks: it scores 74.9% on the SWE-bench Verified benchmark of coding tasks pulled from Github (compared to 74.5% for Claude Opus 4.1 and 59.6% for Google DeepMind’s Gemini 2.5 Pro), it scores 42% on Humanity’s Last Exam benchmark (compared to 44% for Grok 4 Heavy), and scores 89.4% on the GPQA Diamond benchmark, composed of PhD-level science questions (compared to 80.9% for Claude Opus 4.1 and 88.9% for Grok 4 Heavy). The company also says the GPT-5 hallucinates less than the o3 and GPT-4o models; the hallucination measured for GPT-5 was 4.8%, compared to 22% and 20.6% on tests for o3 and GPT-4o. It also claims that the model exhibits less scheming and deceptive behavior than previous models. OpenAI has also stressed the benefits of the model for software development, wishing to facilitate the practice of “vibe coding”. Three APIs – gpt-5, gpt-5-mini, and gpt-5-nano – are being made available to developers.

This opinion article in MIT Technology Review discusses how the most important feature of AI compared to all other technologies, is its ability to self-improve. All AI companies are pushing their AI research to develop models that can develop new ideas. On the one hand, humans may not be able to find cures for cancer or solutions to the climate crisis on their own. On the other hand, self-improving AI can lead to an “intelligence explosion” where AI outperforms humans and gets better at developing weapons and manipulating people. The non-profit group METR is measuring advancements in AI abilities. It says that the time it takes for AI to complete tasks autonomously doubled every seven months since 2019. This figure has shortened to every four months since 2024. The article points to five areas where AI is facilitating improvements.

1. Enhancing productivity. The primary example is software development, thanks to tools like Cursor and Claude Code. The article cites Google CEO Sundar Pichai who says that one quarter of the code at Google is now created using AI.
2. Optimizing infrastructure. AI platforms are traditionally slow, both for training and inference. One application of AI is to design chips that optimize AI performance. The article cites AlphaEvolve which uses the Gemini model to compose algorithms. AlphaEvolve has been used to help run data centers, and Google claims that it saved 0.7% on data center resource consumption. It also designed an OS kernel whose speed improved by 1%.
3. Automating training. Reinforcement learning with human feedback is a common technique for training AI models. The human in the loop nevertheless slows down the process. In the “LLM as a judge” approach, another model replaces the human. Another use of AI is to create synthetic data, as well as creating scenarios for training agentic AI systems.
4. Perfecting agent design. A new area of exploration is using AI to design new AI models. The advent of agentic AI might be an occasion to explore new design alternatives.
5. Advancing research. One quality that humans possess is “research taste” which is the flair to pinpoint promising research questions and directions. This could be a challenge for AI. Nonetheless, the article cites the case of a paper accepted at a workshop at the International Conference on Machine Learning that was entirely developed and written by AI.

This article reports on the cybersecurity conference Black Hat 2025 where agentic AI was mainstream in presented security solutions, notably to cope with the increase in volume of security attacks. The article mentions that cloud intrusions have increased by 136% in the last 6 months, the criminal group Scattered Spider now deploys ransomware in less than 24 hours, and there has been a 220% increase in the number of North Korean operatives posing as remote employees. The security firm CrowdStrike reported that it detected 28 malicious insiders in companies coming from the FAMOUS CHOLLIMA group. This North Korean group is highly dependent on AI to create fake LinkedIn profiles and legitimate credentials, using deepfake technologies to masquerade during job interviews and AI to answer interview questions. Once employed, the operatives use generative AI to do the contracted work while they steal money or IP from the company. They rely on companies in the US and Europe which provide machines to act as relays for remote access.

Among the announcements at the conference were an enhanced version of Microsoft’s Security Copilot which can now correlate threats from several tools, Palo Alto Networks tool Cortex XSOAR has augmented agentic capabilities, and Cisco has released Foundation-sec-8B-Instruct – a conversational AI model developed exclusively for cybersecurity. This model has eight-billion-parameters, outperforms much larger general-purpose models like GPT-4o-mini on security tasks, and can run on a single CPU. Foundation-sec-8B-Instruct has been released with a permissive open-source license, so companies can fine-tune the model for their own environments. All companies interviewed by VentureBeat seem to agree on the need for human threat hunters to accompany AI for their insights and for ways to imagine adversary moves.

One of the challenges to developing large AI models has been the limit of natural and quality data for training the models. Researchers are calling this the “data wall”. Over the past few years, synthetic data is seen as one possibility to break this wall. Synthetic data does not come from the real world, but has the same probability distributions as natural data. SynthLLM is a system for creating synthetic data at scale from a corpus of natural data. It is widely held that the greater the volume of natural data, the better the performance of the trained model. In this article, researchers report on experiments to see whether this scaling law also holds when models are trained on synthetic data. They found that models trained on SynthLLM-generated synthetic data show consistent performance gains for all model sizes. However, performance improvements plateau at around 300 billion tokens; beyond this, adding additional synthetic data does not bring gains. Finally, it seems that larger models need less synthetic data: in tests, an eight-billion-parameter model requires one trillion tokens for optimal performance, while a three-billion-parameter model needed four trillion tokens. The research gives insights into efficiently training models with synthetic data.

The rules of the EU AI Act relating to general-purpose AI (GPAI) models have now come into effect across the Union. A GPAI model is one trained with more than 10²⁰ FLOP of computing power. The AI act’s aim is increased transparency, safety and accountability of models, and the rules focus on clarity on the training data used, better copyright enforcement for training data content, and responsible AI development. The rules apply to all new GPAI models; models released prior to August 2025 must be compliant by August 2027. The EU has released a template to help GPAI providers with their obligations, and to help copyright holders exercise rights. A code of practice has also been developed as a tool to help GPAI providers demonstrate compliance. The code of practice has been signed by numerous Big Tech companies, including Amazon, Anthropic, Google, IBM, Microsoft, OpenAI, Cohere and Mistral AI. xAI signed the Safety and Security Chapter, calling for transparency and respect of copyright, but not the others. Meta has refused to sign the code of practice. Meanwhile, creative groups are unhappy with the act. They say that the opt-out mechanism remains unclear, and judge unfair the principle that AI companies may use copyrighted material unless there is an explicit opt-out by content authors.

This article looks at the emerging trend of “digital resurrection” where AI avatars of deceased people are created to help their loved ones cope with grief. The practice began in China where a “griefbot” or “deathbot” costs less than 3 USD. The industry was valued at over 1.5 million USD in 2022, and is expect to quadruple by the end of 2025. In the West, a YouGov poll suggests that 14% of people would find comfort having an AI avatar of a deceased person. Some experts have criticized the idea. While people preserve memorials and mementos of deceased people, one said that deathbots “get in the way of recognizing and accommodating what has been lost, because you can interact with a deathbot in an ongoing way.”. Another issue is that the deathbot is a sanitized version of a deceased person, as it would hide unpleasant aspects of the deceased’s character. Finally, the issue of permission is not addressed. The deceased person is no longer around to object to being represented in an AI avatar. There have been cases of people in China specifying in their will that they do not want to have an avatar after their passing.