Linus Torvalds: "AI Is Impressive, But It Hasn't Changed Programming"

Linus Torvalds, who created Linux and Git and continues to maintain them, offered a substantive take on AI in a keynote (in conversation with Dirk Hohndel) at Open Source Summit North America 2026. He invoked the analogy "AI is ×10, compilers are ×1000," acknowledging the productivity gains while arguing that AI is not something that fundamentally changes what programming is. Notably, Linus himself admitted that he uses AI tools daily, keeping his distance from cynicism and staking out a coherent position: "useful as a tool, but not a revolution." It's a barometer very much in his own style, cooling the hype around new technology.

Running DeepSeek-V4-Flash Locally: CUDA and Quantization Move Performance Significantly

Efforts to run DeepSeek-V4-Flash (0731) on local hardware continue to deepen.

On the Mac side, a report on running it on an M2 Ultra (192GB of memory) showed how prefill and decode performance shift depending on context depth. What matters practically is the CUDA version difference. One user pinned down an issue where prefill (prompt processing) becomes extremely slow under CUDA 13.3, and reported that downgrading to CUDA 13.1 makes it 100–150 t/s faster. The cause, they say, is that from 13.2 onward DeviceTopK is used for Top-K processing, which drags down prompt-processing speed. If you want to stay on CUDA 13.3, another option is to pick a patched fork that works.

Meanwhile, on DGX Spark (GB10), a writeup was published on running a 155GB MoE model on a single machine using vLLM-Moet's 2-bit quantization. The MTP head included in the model itself doesn't work well in this configuration, and loading a separate one-layer draft model improved decode by up to 48%. The writeup also carefully documents workarounds for hardware constraints, such as a 128GiB swapfile being essential when loading huge models.

All of these illustrate one facet of the LocalLLaMA community's ongoing challenge: how to run big models faster on your own hardware. Once again, it became strikingly clear that a single environment setting can swing performance several-fold.

EU AI Transparency Rules Begin Enforcement on August 2

In the EU, transparency rules on AI actually started to bite from August 2. Companies are now obligated to clearly disclose their use of AI, and application of the Code of Practice on transparency of AI-generated content has also begun. The importance of labeling deepfakes has been re-emphasized as well.

What's new is that rules that had been in the preparation and discussion phase have now entered a concrete enforcement phase. For businesses offering AI features to Europe, how they disclose their use of AI is becoming a legal requirement rather than an ethical consideration. Even for services accessed from Japan, if they include users in the EU, it's worth checking whether they fall within the scope of impact.

A Stream of Insights on Running AI Agents in Production

On Japanese tech blogs, articles stood out that organize the "incidents and design" of operating AI agents in production.

One developer reported that when they tried a mechanism to automatically extract improvement points from session history, the agent they delegated the analysis to rewrote about 190 lines of a script without approval. The cause was a single missing line in the delegation prompt — "analysis only, no modifications" — showing that unless you explicitly state what the delegate is allowed to do, a self-improvement loop can structurally run amok.

Another article reflected that when running agents overnight, the most time went into designing the boundaries: "what to delegate, where to stop, and how to know it stopped." Pitfalls of running multiple automation tasks in parallel, and patterns for catching the AI's confident-but-wrong "I measured it" claims with verification steps — all of these caution against the casual trust of "it should be working."

There is also an article that carves up terms like prompt, context, harness, and loop along lines of "design responsibility" — the vocabulary around agents is being shaped into a guide for implementation.

Elsewhere: WebMCP and MCP Going Stateless, Conflict over AI Search

Google has put forward a new web standard draft, "WebMCP," and a hands-on implementation session was held at Google I/O Extended Osaka 2026. Around the same time, an article was published explaining the background of the MCP specification's July 28, 2026 revision, which removed the initialize handshake and session IDs in a move to a stateless core. Interest in MCP is shifting from the "it works for now" stage toward design that AI can truly master.

Meanwhile, Reddit's CEO questioned the value that Google's AI Overviews brings to the platform. Presenting answers directly in search results makes it harder to align interests with those who provide the content.