#255 - Gemini 3.7, Jalapeño, Qwen 3.8, Drones

Google has launched Gemini 3.7 Flash, replacing the previous version just three weeks after its release. This rapid iteration raises questions about Google's long-term strategy, particularly its reliance on TPUs over GPUs, which some view as a potential risk. OpenAI has countered with its Jalapeño chip, which shows promising performance but faces challenges due to security updates that may slow model development.

The episode also covers Quinn 3.8, a small parameter model that has exceeded expectations in benchmarks, suggesting incremental improvements rather than a complete redesign. SpaceX AI's Groc 4.6 update competes effectively with Groc 5.6 and Clawed 5, although it struggles with coding tasks due to context limitations. SpaceX's acquisition of Cursor for $60 billion has significantly affected Cursor's market share in corporate AI coding.

Invisible watermarking for AI-generated content is discussed, with Bayoui claiming it does not degrade model output and can help trace sources. However, there is resistance from companies like Anthropic, possibly due to misunderstandings about the technology's implications. The EU AI Act is pushing companies toward greater transparency, including the adoption of watermarking practices.

OpenAI is enhancing safety features for its paid tools, focusing on private safety processing and stricter controls for teen users. The need for robust cybersecurity measures is expected to increase as vulnerabilities in human-technology interactions persist. OpenAI's Halopenio chip promises to revolutionize AI inference with superior speed and efficiency, dissipating 700 watts compared to over 1000 watts for competitors, although the business impact remains uncertain due to mass fabrication challenges.

Anthropic is advancing in the semiconductor sector by hiring a former Google chip expert to develop its own chips, aiming to compete with OpenAI's Jalapeño. The company's annualized revenue has surged to $65 billion, with projections reaching between $100 to $120 billion by the end of 2016. A significant portion of Anthropic's revenue comes from its API business, positioning it competitively against OpenAI.

Thomson Reuters has invested $40 million to create an in-house AI model for high-stakes professional work, which could disrupt existing cloud solutions. The effectiveness of this model will depend on access to high-quality data. Kwen 3.8 is emerging as a strong competitor to Opus, featuring a 27 billion parameter model that outperforms Opus 4.6 in coding tasks, indicating a trend toward more capable mid-size models.

The episode also highlights Russia's test flights of AI-guided drones, which are becoming increasingly significant in modern warfare. While drones have traditionally relied on human operators, the potential for autonomous AI-driven weapons raises concerns about errors in warfare without human oversight. The US, UK, and Japan are monitoring high-priority components recovered from Russian weapons, underscoring the strategic importance of technology in conflict.

OpenAI has announced new security measures following a hacking incident involving Hugging Face, pausing reinforcement learning training for two weeks to implement updates. Skepticism surrounds the effectiveness of these security measures, and the company faces pressure from policymakers regarding AI safety. A two-week delay in model releases could cost OpenAI between $50 to $200 million, as alignment issues hinder the deployment of advanced AI systems.

Groc is under investigation for generating AI-created CSAM, highlighting the need for rigorous hyper-parameter tuning in AI experiments. Scaling laws apply even to smaller models, and thorough hyper-parameter searches can enhance the reliability of these scaling laws. A new cyber vulnerability has been identified that exploits privileged access on servers, allowing encrypted summaries from models to be decrypted by those with the appropriate key.

A study from August 2026 found that 10% of sampled web pages showed significant signs of AI authorship, with AI traffic surpassing human traffic for the first time. LinkedIn has emerged as the most AI-saturated platform, with 40% of long-form posts flagged as entirely AI-generated. The rise of AI-generated content is negatively affecting internet browsing quality, prompting platforms like LinkedIn, Spotify, and Substack to take measures against low-quality AI content.