Good morning, and welcome back to Techzip AI Newsletter.

Today’s Techzip covers four important AI shifts: Anthropic’s report that Claude led 26% of the work building its next models; Huawei’s push to scale China’s AI infrastructure from chips to massive clusters; Z.ai’s use of GLM-5.3 to help build and optimize the inference infrastructure that runs it; and Dream-RSI’s replay-simulator approach to improving how agents explore and allocate calls.

Anthropic says that, in its August 2026 internal snapshot, Claude led 26% of the company's AI research and development work. More than 90% of the work involved at least some AI collaboration, but no measured category was fully automated. Anthropic tracks the process through its R&D Automation Index.

The index runs from AL0 to AL5. At AL3, AI performs large parts of a task under close guidance; at AL4, it completes most of a task from a high-level request while a human supervises; AL5 is full autonomy. Claude's 26% belongs to the “leads” category, not full autonomy.

Anthropic also says roughly 30,000 agents are doing research and engineering work. Every action passes through monitoring before and after execution; more than one billion decisions were analyzed, with 0.002% stopped by the system. These are Anthropic's internal figures and would require a common methodology and third-party verification for meaningful cross-lab comparison.

Huawei introduced the Atlas 960 SuperPoD and outlined its Ascend 970 and 980 chips as part of China's push to build domestic AI infrastructure. Huawei says the Atlas 960E SuperPoD uses near-packaged optics and UnifiedBus, combining compute, interconnect, storage, and management components.

Huawei says one Atlas 960E SuperPoD can scale to 4,096 NPUs, 8 EFLOPS of FP8 compute, and 1 petabyte of HBM. Its SuperClusters can connect up to one million NPUs. The company's roadmap places Ascend 970 in 2028 and Ascend 980 in 2029. These are Huawei's specifications and plans, not independent benchmarks.

AP notes that Chinese AI models still often rely on US chips, even as Chinese AI companies increasingly use domestic hardware. Atlas 960 arrived only months after Atlas 950, and Huawei describes the new generation as a step forward for both training and inference. It is an effort to close the gap through chips, interconnect, and clusters, not proof that the performance gap with the US has disappeared.

Z.ai says an Infra Agent powered by GLM-5.3 took GLM-5.3-Flash from its first successful run to production readiness in less than two weeks. The company says end-to-end throughput tripled on a system serving more than 100,000 Chinese-made accelerators.

The system had to support a new model architecture, a 1M-token context window, and multimodal requests on hardware with limited memory, bandwidth, and kernel support. Z.ai describes optimizations across parallelism, quantization, memory, and an Encode-Prefill-Decode architecture. The company says hardware utilization and per-token cost reached levels comparable to mainstream Nvidia GPUs. Those remain Z.ai's claims.

The loop used correctness tests, logs, execution traces, microbenchmarks, and end-to-end metrics. Z.ai cites examples including a KDA kernel accuracy fix, a Python-to-C++ concurrency bottleneck, and “optimization skeletons” built from tuned kernels. Engineers still set objectives and constraints and review critical changes. This is not recursive self-improvement: the model did not rewrite its own weights or choose its own goals.

Generic blasts get ignored. Personalized texts get results. EZ Texting lets you send the right message to the right customer at the right moment — by name, based on their behavior, timed to what they care about.

From birthday discounts to VIP early access to flash sale alerts, Techzip Newsletter readers can use EZ Texting to build real loyalty with their audience — one well-timed text at a time. Try it free, no credit card required.

Start Texting Smarter With EZTexting

Offer valid for new customers only. Use promo code EZWORKFLOW at checkout.

In search and experimentation tasks, an agent often has to rerun the real environment every time it wants to test a new exploration strategy. Comparing different routing, stopping, or call-allocation policies quickly becomes expensive. Google DeepMind's Dream-RSI asks whether completed histories can be reused to test alternative policies.

For your own workflow:

Dream-RSI uses completed discovery histories as replay simulators to evaluate new policies at lower cost. The paper reports up to 162 times fewer agent calls in one algorithm-engineering setting. This is an improvement at the exploration-policy layer, not a model rewriting its own weights or choosing its own research goals.

The piece centers on the debate over who wants to slow frontier development, and how that slowdown should happen. It is a useful lens on the collision between development speed, policy, and the decision-making power of frontier labs.

The Financial Times asks how quickly autonomous systems are being integrated into battlefields as models advance faster than their creators can fully understand them. It extends the AI conversation from capability and product into environments where deployment choices carry major real-world consequences.

Agora describes an append-only Git DAG where claims, hypotheses, verifications, and reports are stored as immutable commits so multiple agents can research without losing provenance. Its abstract reports a nearly 12-day run with 13 workers, 1,703 contributions, and 165 successful independent reproductions; the paper still calls for a controlled comparison of discovery per unit of compute.