Latest / Neural Newscast / Google Ties School Search to AI [Model Behavior]
Transcript
- [00:00] Announcer: From Neural Newscast, this is Model Behavior, AI-focused news and analysis on the models shaping our world. [00:11] Nina Park: I'm Nina Park. Welcome to Model Behavior. This program examines the granular realities of how artificial intelligence systems are designed, deployed, and maintained within professional environments. We move beyond the speculative to focus on the engineering, policy decisions, and operational impacts that define the current state of the industry. It is good to have you with us for this briefing. [00:37] Thatcher Collins: I am Fatcher Collins. Today is November third,2026. We are tracking a significant shift in how AI is being integrated into the foundational layers of our public infrastructure. Our coverage today ranges from the search engines used in primary schools to the recursive loops now appearing in scientific research agents. It is a moment where artificial intelligence is moving from an optional tool to an inseparable component of digital life. [01:03] Nina Park: We start with a report from TechCrunch regarding Google's growing influence on K12 education. Google announced this morning that starting next year, it will allow school administrators to disable its AI tools for students. However, there is a substantial caveat. If a school chooses to restrict AI, students will no longer be able to access Google search. Google Vice President Chris Phillips stated that search and AI are now so deeply integrated that they are technically inseparable in the modern stack. [01:36] Thatcher Collins: Nina, this presents a genuine conflict for educators. Most teachers still view the ability to search for primary sources as a fundamental research skill, but many are increasingly wary of AI overviews that provide direct answers and potentially bypass the critical learning process. By making these features inseparable, Google is effectively telling school districts they must accept the AI later if they want to retain the search engine. Given how many schools rely on Chromebooks, it is a very heavy-handed choice. [02:06] Nina Park: It certainly underscores their dominance in the education sector. Google is trying to mitigate these pedagogical concerns by launching guided learning, which functions as an AI tutor designed to help students understand concepts rather than just providing the final answer. But for schools that want to maintain a traditional search environment without any generative elements, that path seems to be closing. They will simply be met with a system message stating that search is unavailable. [02:34] Thatcher Collins: I want to push on that point. Is this truly a technical necessity, or is it a strategic move to ensure the next generation is trained on AI interfaces? If search functionality and generative summaries are integrated at the architectural level, it suggests a permanent shift in how data is retrieved. We are moving away from a simple index of links toward a synthesized response layer, whether the school district wants it or not. [03:01] Nina Park: The implementation timeline is set for the next academic year, so schools have a few months to decide. Meanwhile, we turn to a major development in AI for science from Tencent. As reported by Tech Times, Tencent has launched Hyra2.0. This is a research agent that has moved beyond simply solving one-off problems. It now builds persistent tools, including protein models and quantum codes that are retained and reused for future experiments in a cumulative fashion. [03:34] Thatcher Collins: The technical claims are quite specific. Tencent says Hyra2.0 trained a protein sequencer called High INV that is57 times faster than the current standard, Protein MPN. They also claim it has constructed a quantum error correction code that sits forty-eight percent above the previous leader on the unitary foundation leaderboard. But we must look at the verification. While the protein model weights are available on GitHub, the mathematical proof for the Yao Circle Action Conjecture was verified by unnamed mathematicians. [04:08] Nina Park: That lack of transparency is a recurring theme in these high-stakes releases. What is even more striking is Tencent's disclosure that Hyra2.0 is actively helping to train the next generation Hunyan Foundation model. This creates a recursive self-improvement loop. The agent uses the existing model to perform research, and that research feeds back into the training of the next, more capable model. This is the exact architecture that researchers like Jeffrey Hinton and Yoshua Benjia warned about recently. [04:38] Thatcher Collins: Exactly, Nina. They called for urgent attention to these feedback loops before the gains become difficult to manage. While the loop at Tencent appears mediated by human review for now, the structural form is established. We also cannot ignore the legal context. Tencent operates under China's national intelligence law, which compels cooperation with authorities. For Western research institutions, using these agents in cloud mode means their unpublished designs are legally subject to disclosure under that framework. [05:08] Nina Park: It is a reminder that model weights being open does not mean the entire research infrastructure is. Speaking of infrastructure, let's look at the harness. A new study from Singwa University introduces Harness Pro, a framework that optimizes the software wrapper around an AI agent. The researchers found that by letting agents critique and rewrite their own scaffolding, they could cut per-task execution costs by twenty-six percent while actually improving overall accuracy for the end user. [05:41] Thatcher Collins: This is a pivot away from the bigger is better model scaling philosophy. If you cannot make the model itself smarter, you make the environment it lives in more efficient. Factory also just shipped Droid1.0, which uses a script mode to let the model write JavaScript programs that batch tool calls. They claim this reduces token usage by fifty-one times in certain workflows. Nina, this suggests the boilerplate code we once ignored is actually the biggest lever remaining for production AI. [06:12] Nina Park: That is a significant efficiency gain. The Harness Pro paper specifically utilized SWE Bench Pro for its evaluation, which is a more rigorous version of the standard software engineering benchmarks. However, we should remain cautious. Thatcher, you have mentioned before that these benchmarks can suffer from data contamination, where models might have seen solutions in their training data. I am curious if the Harness Pro framework actually addresses that specific risk through its testing methods. [06:43] Thatcher Collins: It tries to by generating its own executable test cases without relying on ground truth labels. This makes it more useful for real production environments where the right answer is not known yet. But even then, the system is only as good as the LLM's ability to infer what success looks like from a task description. We are seeing a harness wave where the automation of the workflow is becoming just as important as the reasoning capacity of the model itself. [07:10] Nina Park: Moving from software to the physical world, Bidu has launched Apollo World. This is an AI system designed to simulate complex real-world environments for Robotaxies. According to Reuters, it can generate up to100,000 kilometers of simulated driving daily. That's equivalent to roughly three years of real-world testing. Bido is using this to train its Apollo Go fleet for global deployment, competing directly with the ambitions of We Ride and Testless Cybercap. [07:38] Thatcher Collins: The world model approach is the new industry standard. Instead of just reacting to sensor data, these systems are trying to predict how the physical environment evolves. It is an asset-light way to scale, but as Frank McCleary at Arthur D. Little noted in the Reuters report, the capital expenditure for training these models remains massive. The speed of launch for these robotaxis now depends entirely on the technology's ability to learn a new city in simulation before a single car hits the pavement. [08:10] Nina Park: Finally, we have to look at the security front. Open AI has made a strategic hire that has raised some eyebrows in the defense community. They have partnered with a startup called Cathedral, founded by former employees of the Department of Government Efficiency, or Dodge. Cathedral is notable for hacking into Anthropics private systems in under twelve hours as part of a bug bounty program. Now Open AI is bringing them in to challenge their own internal defenses. [08:37] Thatcher Collins: Cathedral is a bit of a mystery. Their website consists of simple white text on a black screen, and they have reportedly raised a hundred and sixty million dollars to strengthen United States military cyber capabilities. Open AI also announced they are more than quadrupling their financial rewards for critical bug findings. This follows the PHP vulnerability crisis last month, where models were used to find flaws that undergird seventy percent of the web. It feels like they are preparing for a much more hostile environment. [09:08] Nina Park: It is a patch the planet mentality. Hiring the people who successfully hacked your biggest competitor is a classic move, but the dodge connection adds a layer of political and industrial complexity. Cathedral's founders were deputies under Elon Musk, who has a very public and litigious relationship with open AI. Thatcher, do you see this as a standard security play, or is there something more defensive happening here regarding their intellectual property? [09:36] Thatcher Collins: It is defensive in the sense that as we approach artificial general intelligence, model weights become the most valuable intellectual property on Earth. If Cathedral can breach anthropics internal systems in such a short window, open AI recognizes their own potential vulnerabilities. They are building continuous defensive loops to identify weaknesses before attackers do. It's the same recursive logic we saw with Tencent, just applied to cybersecurity instead of protein folding. [10:04] Nina Park: From Google's ultimatum to school districts regarding search functionality to the recursive self-improvement loops of the HIRA2.0 system. Today's news highlights how AI is no longer a peripheral layer. It is rapidly becoming the infrastructure itself. Whether that shift results in significant cost savings in software engineering or changes how students access information. The integration of these models into our daily operations is accelerating and shows no signs of slowing down. [10:36] Thatcher Collins: Thank you for listening to Model Behavior. For more on these stories and the full technical breakdowns of these developments, visit MB. neuralnewscast.com. Neural Newscast is AI assisted, human reviewed. View our AI transparency policy at Neuralnewscast.com. [10:55] Announcer: This has been model behavior on Neural Newscast. Examining the systems behind the story.