On August 3, 2026, the AI industry was abuzz with intensifying model development competition, seismic shifts in hardware, and debate over how companies should judge their AI investments. While OpenAI revealed the name of its next flagship model for the first time, xAI and Chinese manufacturers are lining up their challengers. Here are the key points of the day.

OpenAI Reveals the Name of Its Next Model “Astra”, with New Results on 10 Unsolved Problems

OpenAI announced that, using an internal version of its next flagship model “Astra”, it has obtained new results on 10 unsolved problems in mathematics and theoretical computer science. This appears to be the first time the “Astra” name has been made public. Compute costs were kept to around $2,000 by the company's “Sol” metric, and formal proofs produced with the Lean proof assistant have reportedly been published on GitHub. The move signals ambitions for serious application to advanced mathematical research, and stands as a milestone showing what reasoning models can achieve.

On August 1, reports had described “Astra” as a new family of models specialized for long-horizon reasoning; this announcement fills in the concrete substance — results on unsolved problems backed by formal proofs. That said, whether the 10 problems count as “solved” in the strict sense awaits expert verification.

Industry Watch: Grok 4.7 Nears Arrival, Microsoft's “Polaris” May Take Over Copilot

AI Daily Digest rounded up several industry developments from the same day. According to the report, the arrival of xAI's next model “Grok 4.7” is drawing near, and Microsoft's “Polaris” appears to be moving toward taking over Copilot. There is also speculation that OpenAI's IPO has been postponed.

Definitive information is limited on all of these fronts, so treating them as industry speculation is the safest course for now. Still, the fact that a reshuffle of the frontier model camp and turbulence in capital markets are being reported at the same time suggests the AI business is reaching something of a turning point.

China's DFSX Touts Twice the Memory Bandwidth of NVIDIA's GB200

China's new AI system “DFSX” has reportedly achieved twice the memory bandwidth of NVIDIA's GB200 NVL72. According to reports, it features 14nm-process “supernodes” and a design that omits microbumps to vertically stack compute units and memory. Memory bandwidth is often the bottleneck in LLM inference, so how much real-world effect this architectural ingenuity delivers will be worth watching.

More on Local Optimization of DeepSeek V4-Flash: Be Careful with Quantization, 16 tok/s on Mac

In the local LLM community, optimization of DeepSeek V4-Flash continues to be hotly debated.

First, quantization needs careful handling. Community testing found that quantizing the KV Cache from BF16 to Q8 produced a measurable impact on quality in DeepSeek V4-Flash (worse PPL and KL divergence). Under the same conditions, the impact on Qwen 397B was reportedly small — a fresh reminder that quantization tolerance differs from model to model.

On the speed side, there is good news too. In a case running the IQ3XXS quantization on an M1 Ultra (128GB), speed improved from 5–6 tok/s to 15–16 tok/s after engine fixes, with output quality also reportedly improving.

Also making the rounds: the llama.cpp team has released the official “llama.app”, complete with a DMG-based installer for Mac and a “llama serve” command that starts without any arguments. The ease of use that tools like Ollama have long provided is finding its way into llama.cpp itself.

The Front Lines of AI Adoption at Japanese Companies: The Battle over Cost, ROI, and Organizational Culture

In Japan, figuring out how to run AI with a solid return on investment has become a practical management challenge.

Bizreach uses AI at a scale that consumes roughly 10 billion tokens per month, and how it evaluates that ROI amid growing cost concerns is drawing attention. CyberAgent is holding an AI contest with a 10-million-yen prize, deliberately waiving any “feasibility” requirement as a device to change how its employees think.

Some companies are showing results in hard numbers. Salesforce says that through its own internal practice (“Customer Zero”) it completed 4.31 million customer interactions with AI and increased its number of business negotiations by 40–50%. The company cites choosing “areas where data and KPIs are in place” and delegating simple, high-frequency tasks first as its iron rules for success.

Workplace adoption is uneven, too. In a Ragzas survey, engineers and IT staff did not rank first in generative AI usage by occupation — a different job type came out on top. It's the structure — tools deliver no results if nobody uses them — showing up as a number.

Note that AI has become a threat to defenders as well. The arrival of frontier AI has completely changed the nature of web attacks, with cases reported of attacks evading WAFs 89% of the time. The need for preventive measures against faster, automated attacks is being pointed out once again.