Gemini 3.5 Flash's "Hidden Cost" — Running Costs Reach 5.5x Its Predecessor's

Announced at Google I/O 2026, Gemini 3.5 Flash delivers major benchmark gains over its predecessor — but behind the scenes, its running costs have surged roughly 5.5x. On agentic tasks in particular, the model requires more interaction steps than competing alternatives, reportedly incurring as much as 75% in additional costs — more than even the pricier Gemini 3.1 Pro.

And this isn't a Gemini-only problem. Anthropic and OpenAI have also raised prices on their latest models, and as the industry tries to recoup massive investments, the "higher performance = higher cost" trend is accelerating across the AI sector. For developers who have benefited from cheap APIs, it's time to rethink their cost assumptions.

Training a Billion Gaussians on a Single GPU — TideGS Breaks the Scale Limit of 3D Reconstruction

Training 3D Gaussian Splatting (3DGS) hits a GPU memory bottleneck as the number of primitives grows, with tens of millions previously being the practical ceiling. TideGS, featured on Hugging Face Daily Papers, is an out-of-core optimization framework that overturns this limit at its root.

The key to TideGS is exploiting the fact that 3DGS training is inherently "sparse": only the Gaussians visible from the current camera batch are active in each iteration. It's a conceptual shift that treats GPU memory as a working-set cache. Parameters are managed in a three-tier SSD–CPU–GPU hierarchy, combining three techniques: block virtualization, asynchronous pipelining, and trajectory-adaptive differential streaming.

In experiments, the team trained over one billion Gaussians on a single 24GB GPU and surpassed existing single-GPU methods in reconstruction quality on large-scale scenes. That's a two-order-of-magnitude scale-up over standard in-memory training, which had been capped at around 11 million Gaussians.

Gartner Forecast: AI Shifts to Growing Employment from 2028

According to a forecast from Gartner's HR research division, from 2028 onward the jobs AI creates will outnumber the jobs it eliminates. While job displacement will progress in the short term, the long-term outlook is that new AI-related roles and organizational expansion driven by productivity gains will lift overall employment.

This offers one perspective on the debate over AI's impact on employment — but of course, "what kinds of jobs grow" becomes the critical question. Beyond simple headcount changes, the effects on wage levels and the quality of work will also demand close attention.

From the AI Coding Trenches — 100K Lines of Rust and "AI-Ready" Monorepos

On Hacker News, an article summarizing the experience of writing 100,000 lines of Rust with Claude Code and Codex has drawn significant attention. It presents spec-driven development and contract-based design as the keys to improving AI coding accuracy.

Meanwhile, an "Ask HN" thread discussed how to make large monorepos work well with AI tools. The prevailing view: AI readiness ultimately comes down to the same thing as code that's readable for humans — making patterns explicit through CLAUDE.md files and cleaning up legacy patterns proved effective.

New Trends in AI Code Review

JetBrains has added a "Self-Review with AI" feature to IntelliJ IDEA, and the experience of having AI automatically review your code before commit is generating buzz. It's an approach that lightens the burden of manual self-review while also saving reviewers' time.

On Japan's Qiita, a skill configuration for rigorous PR reviews using Claude Code has been published. While performance improvements since Opus 4.5 have made large-scale changes easier, growing concerns about declining implementation understanding are fueling a movement to have AI itself take on the quality gate role.

---