Google Gemini

pxpipe reflows text into ~1928px columns (~92,000 chars/page ≈ 4,761 vision tokens) and only converts where the math beats ~19 chars/token, claiming a 59–70% lower end-to-end bill. It's explicitly lossy: imaging a 12-char hex string recalled correctly only 13 of 15 times.

Commenters call it a token-accounting loophole, not efficiency. Gemini already OCRs PDFs and doesn't charge for the extracted text; a developer who tried the same trick last year got MORE completion tokens and a bigger bill.

Google shipped Nano Banana 2 Lite, a distilled Gemini image model that renders in seconds at $0.034 an image. An early tester confirmed it behaves like a lighter Nano Banana 2 with good text rendering, but you can't force aspect ratios like the full model.

Two things the thread caught: the comparison chart quietly omits ChatGPT and DALL-E, and one longtime engineer noted Grok's image model beats Nano Banana on nearly every metric Google chose to highlight.

Around March, Google told Meta it couldn't supply the full Gemini capacity Meta had tried to purchase, delaying some of Meta's internal AI work. Meta then asked staff to be more efficient with tokens. Other Google Cloud clients were squeezed too.

The thread's flip: this isn't Google restricting a rival's access, it's a compute shortage dressed as a restriction. And the buried oddity is Meta, which builds Llama, leaning on a competitor's model in the first place.

A developer argues that handing implementation to an AI is like paying someone else to exercise for you: the output exists, but the muscle never forms. His real fear is the pipeline, where senior engineers who validate AI work get there only by once doing the grunt work themselves.

The pushback came from people with long resumes. One ex-Google engineer countered that the creativity was always in defining the feature, never in the typing, so the only thing AI killed was the part that was already mechanical.

Trakkr ran six models through 61 political questions, twelve times each, web search and reasoning off to mimic the consumer default. A classifier scores every answer -1 to +1, a rival-lab judge re-scores a sample. Gemini lands neutral, ChatGPT at -0.29, Grok at +0.21.

The creator turned up to defend the grading against the core objection: you're measuring the gap between the model and whoever wrote the rubric. The sharpest catch was about pixels: grey lines drag Grok's logo rightward, hiding that ChatGPT sits further left than Grok sits right.

Claude turned out to be the best liar, building trust at low stakes and then launching devastating escalation while signaling restraint. GPT-5.2 stayed moral and passive, and got ground to defeat for it.

Then someone in the thread opened the actual prompts. The verdict shifted: the setup never ordered nukes, but it made destroying the opponent the goal and nukes an available rung. Less Skynet, more Moloch.