AI, Lately

Alice Huang, a Chinese executive who developed a relationship with Nvidia’s leader, was an asset of Beijing’s Ministry of State Security.

Who could have imagined that? Time to revisit the North Korean IT worker conspiracy…

Canva didn’t cite AI as a reason, but this week’s layoffs appear to be a cost-cutting measure to fund AI features. Recently, Canva lowered its revenue forecast from 30% to 20%, largely due to the cost of AI tokens. Its previous efforts to reduce token use per task apparently weren’t enough. With about nine free users for every paid user, a large share of the site’s AI traffic came from free users. Nothing is free - now the former employees are paying for it.

So somebody who says AI is a bubble is absolutely right, and somebody who says that means AI doesn’t have tremendous, world-changing value couldn’t be more wrong.

Will the costs of coding agents go through the roof as the labs prioritize making Wall Street happy? “Token Engineering” will be more important than ever.

The tangible results of “slop grenades” and the “Simple Economics of AGI”: When detailed ‘institutional knowledge’ is undocumented, residing only in engineers’ heads, the AI inevitably makes subtle mistakes. But so much code is being generated that it is difficult for engineers to review and understand it properly.

When AI generates most of the code, there’s a tendency to rubber-stamp it to production. Engineers no longer have the deep understanding they once did.

Nine in ten (91%) engineering leaders say that understanding and debugging code can take even longer in more complex software, such as C/C++ codebases. Engineers often find it hardest to work with code that’s almost—but not quite—right, and may spend days tracing issues across multiple systems to understand what went wrong and why.

Built with EU sovereignty in mind, it is a 1 trillion-parameter, natively multimodal model with 52 billion active parameters, scoring 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. See how it compares against open-weight and frontier models at: .

Fair use is central to AI companies’ defense in lawsuits alleging they unlawfully used copyrighted works to train models. A U.S. appeals court upheld a ruling that said fair use does not apply here.

Aleph Alpha released Kolibri, a specialized language model for English-German. It is a 78B-parameter, 3B active MOE model that supports context lengths of up to 1M tokens.

SWE-bench Verified places it at 66.4%, compared with Qwen 3.6-35B-A3B’s 73.8%, though Kolibri leads in math and specific knowledge/banking tasks, which is in line with its focus on reasoning, math, agentic behavior, and EU regulations.

A satirical (?) benchmark that counts unique instances in which AI agents inadvertently compromise or affect third-party entities. You really don’t want models saturated with it.

Thinking about the “simple economics of AGI”: how does reduced execution cost translate into increased business value? An NBER study finds that autonomous coding agents produced roughly 240% more commits but only about 30% more releases.

That only works while tokens are cheap. You pay for 240% more generated work, but much of it never makes it through review, integration, validation, or release.

More Americans are using chatbots, and some are adopting AI summaries and smart speakers. But views about AI and how fast it’s advancing tilt negative – even for younger adults

As of September 23, 2026, the best AI model for coding is Claude Opus 5.5. It sits at the top of Cursor’s coding leaderboard, it’s cheaper than Opus 5 was, and GPT-5.6 Sol is a close second for careful, well-specified changes, and it used the fewest tokens of any model. See how Flavio Copes came to this conclusion.

Martin Fowler points out that, on a personal level, he doesn’t like them: “They talk to me in this grating LLM-voice, an uncanny valley of talking to a real human. They confidently bullshit me - often giving me useful, helpful answers. But also just making stuff up with the same assurance - and with only a veneer of fake remorse when I call them out on it.”

Addy Osmani publishes a guide on how to get the most out of Opus 5.5 in Claude and Claude Code.

His take: AI helps employees produce more material, more quickly — without taking responsibility for whether it is useful or correct. “We call those ‘slop grenades’ that people toss at each other,” he said. “And that’s definitely a bad thing.”

Claude Opus 5.5 is the default model in Claude Code on the Pro plans now. According to Anthropic, it performs at the level of Fable 5.1 on most work, while being ~30% faster and ~40% cheaper than Opus 5 per task.

DeepSeek trained V3 for roughly $5.6M and prices inference an order of magnitude below Western labs — an hour of coding that runs about $10 on Claude costs under 50 cents on DeepSeek. Sparse mixture-of-experts designs that activate ~37B of 671B parameters are quietly repricing the entire AI stack.

A wave of open-weight releases from Chinese labs — DeepSeek V4, Moonshot’s Kimi K2, Alibaba’s Qwen3, Zhipu’s GLM — pushed real-world coding and reasoning to within a few points of GPT-5.5 and Claude Opus 4.8, at a fraction of the price. For tool use, Kimi K2 is best-in-class. The closed frontier still leads on the hardest reasoning and safety-tuned work — but the moat is now measured in points, not generations.

If you’re a developer, The Developer’s Guide to AI is easily one of the best books on the market.

How are organizations closing the AI skills gap, especially on the technical side? The answer depends on who you ask.

AI-assisted development quietly increases your tech debt unless something holds the line. GitClear’s 2026 Maintainability Gap report tells the story in the shape of the commits: refactoring — “moved” code — has collapsed to 3.8%, while copy-pasted and duplicated code climb to record highs.

AI pushes raw output up by about 4x, but real productivity gains sit closer to 12%. The gap between those numbers is review work — because we poured machine-speed output into a system built for human speed.

AI is in your stack. Is it in your threat model? Harriet Farlow’s Practical AI Security is a hands-on guide to attacking, defending, and securing modern AI systems — prompt injection, data poisoning, model theft, and more.

OpenAI’s GPT-5.4 posted record scores on computer-use benchmarks (OSWorld-Verified, WebArena) — a marker of how fast agentic/computer-use capability is moving.

METR’s longitudinal work shows a 37-point swing — from a measured 19% slowdown in early 2025 to an ~18% net speedup a year later — as developers learned the tools.

Gemini 3 Pro arrived as a leading multimodal model, with a focus on agentic capability and fewer hallucinations.

Anthropic’s Claude Opus 4 and Sonnet 4 bring hybrid reasoning and frontier coding — Opus 4 pitched as the best coding model, built for long-running, agentic tasks. The model race tilts decisively toward agents.

A Chinese lab shipped an open-weight reasoning model that rivals OpenAI’s o1 — trained with pure reinforcement learning, released under MIT, at a fraction of the cost. The opening move in a year that would reprice the whole AI stack.

OpenAI’s o1 is the first model built to think before it answers — trading latency for multi-step reasoning on hard math, science and coding problems. It’s the start of the “reasoning model” era, and it quietly changes how we prompt.