If there is one topic that has been absolutely unavoidable during the last couple of weeks, it is the aftermath of the GPT-5 launch. Not only did GPT-5 fail to meet the overly hyped expectations that built up in the lead-up to the model's launch, but its rollout also exposed how little OpenAI and other tech companies understand the bonds forming between chatbots and their users.

Shortly after OpenAI CEO Sam Altman described GPT-5 as "a PhD-level expert in any area you need", examples where the model noticeably missed the mark started popping up all over social media platforms, including the apparent revenge of the "rs in strawberry" problem, as users showed how the new model struggled to provide a correct answer when prompted to count the bs in "blueberry". Other popular examples included prompting the chatbot to generate labeled maps of the US and timelines of the most recent US presidents. In addition to some ChatGPT outputs being genuinely hilarious, the trend re-sparked the debate on what it means for an AI model to be intelligent, and on the adequacy of the current measures of model intelligence.

Then, on top of the constant stream of instances of not-quite-PhD intelligence, OpenAI's decision to sunset all its previous models sparked unexpected outrage among users, with the common complaint being that the new model was not friendly enough. This, of course, is a broad generalization of the issue, given that some users were clearly in distress and grieving the loss of a close companion. OpenAI eventually brought back some models, including 40, for paying subscribers and acknowledged it would offer its users sufficient notice if it decided to follow through with the model deprecation. The company also rolled out an update that it claims will make GPT-5's personality warmer and more approachable.

More generally, and perhaps worryingly, some have raised concerns that GPT-5 not living up to the expectations is a sign that generative AI's current strategy, which depends entirely on resource scaling, may indeed be reaching its limits. It is, of course, too early to tell what the future will bring. However, the enthusiasm surrounding the promise of imminent artificial general intelligence has noticeably cooled down, with even Altman himself trying to walk back some of his previous claims about the imminent and inevitable arrival of AGI, and recognizing that generative AI may indeed be in a bubble.

Other noteworthy recent headlines

Multiverse debuts tiny AI models that run offline without compromising performance: Spanish AI startup Multiverse Computing released two ultra-small AI models called SuperFly (94M parameters) and ChickBrain (3.2B parameters) that can run offline on devices from smartphones to IoT appliances while matching or exceeding the performance of their larger parent models.

Estonia-based Better Medicine raises €1M for its AI-powered detection tool for kidney cancer: Estonian medtech startup Better Medicine raised a pre-seed €1M round to expand the EU rollout of its CE-certified AI kidney cancer detection tool and prepare for U.S. market entry. Better Medicine aims to address critical radiologist shortages with its AI-powered lesion detection solutions.

Ai2 releases MolmoAct: an open-source AI model that "thinks" in 3D space: Ai2 released MolmoAct, the first open-source Action Reasoning Model that can "think" in 3D space for robotics applications, achieving state-of-the-art performance while being trained more efficiently than competing models from major tech companies.

Humanloop's team joins Anthropic as competition in the enterprise AI market tightens: Anthropic acquired the Humanloop team, including its three co-founders and a dozen engineers, in an acqui-hire to strengthen its enterprise AI strategy with expertise in prompt management, LLM evaluation, and AI observability tools.

Cohere secures $500M at $6.8B valuation, adds Joelle Pineau as Chief AI Officer: Cohere raised $500 million at a $6.8 billion valuation to accelerate its security-first enterprise AI platform, adding former Meta AI research head Joelle Pineau as Chief AI Officer and former Uber acting CFO Francois Chadwick as CFO.

Former Google engineer launches Continua with $8M to transform group chats with AI: Continua, founded by former Google Distinguished Engineer David Petrou, raised $8 million in seed funding to develop AI agents that can naturally participate in group chats on SMS, iMessage, and Discord, helping coordinate tasks and reduce conversation chaos.

Seoul's Datumo Raises $15.5M Series B to Challenge Scale AI with AI Safety Focus: Seoul-based AI startup Datumo raised $15.5 million in Series B funding led by Salesforce Ventures to compete with Scale AI. Datumo aims to differentiate itself by providing automated AI model evaluation and safety tools for enterprises, and leveraging its unique licensed datasets.

NVIDIA unveils new world models and AI infrastructure tailored for robotics applications: NVIDIA unveiled a robotics development suite at SIGGRAPH 2025, featuring the Cosmos Reason model and other advanced simulation tools. Major companies are already adopting NVIDIA's technologies to accelerate the development of physical AI applications across trillion-dollar industries.

Ukraine Launches AI Factory to Secure Digital Sovereignty: Ukraine has launched the AI Factory initiative to build domestic AI infrastructure and develop a national large language model, aiming to achieve digital sovereignty and become a top-three AI nation by 2030.

ElevenLabs launches an AI music generator that creates full songs from text prompts: ElevenLabs' Eleven Music is an AI music generation service that has reportedly been trained on properly licensed data. However, its commercial viability remains uncertain due to its restrictive terms of service, which contradict some public claims about commercial use permissions.

DeepMind's Genie 3: A real-time world model for training AI agents: Genie 3 is a real-time world model that generates 3D environments that remain coherent for minutes at a time and can be customized using "promptable world events. DeepMind has highlighted Genie 3's potential for creating new training and education opportunities for a wide variety of agents.