sourweasel

I noticed this recently with the user submitted benchmarks on Kaggle. For these obscure tests that the models haven't seen, Gemini 3.8 is often on par or beating other frontier models. Gemini is a bit shite at the game checkers though, for some weird reason it performs poorly on those benchmarks.

I would recommend adding the MiniCPM5-1B model. Surprisingly coherent for a 1B model and performs well. On my pixel 9 I get 33 tok/s on CPU. On GPU I get about 26 tok/s but prefill jumps to nearly 500.

I've noticed that some companies have flat out dropped the high memory configurations from their lineup that were offered just months ago at inflated prices. OnePlayerX had a 128GB Strix Halo tablet, RedMagic had a 24GB Android phone - both have been removed from their sites. It's a similar situation for the new Thinkpad lineup.

Same experience for me. I get endless Cloudflare loops for many sites due to using less common browsers. Given that the goal is is distinguishing bots from actual humans, it seems to be a rather high failure rate. At least I think I'm human...

One method that does seem to help is being sloppy when clicking the "I am a human" checkbox. I suspected that bots were more likely to perfectly click the center of the checkbox every time and decided to try clicking the padding around the checkbox instead (which still registers as a click). It seems to be more successful, but it could be my imagination.

The first sentence of the main site is "D2: Declarative Diagramming - A modern language that turns text to diagrams", followed by multiple code snippets and the resulting diagrams.

My favorite so far:

>”…in 2015, 1.2 million people kicked the bucket from kidney dissatisfaction, a development of 32% since…”

Are you manually rendering previews of its OpenSCAD output to create images for it to review, or do you have a workflow that automates that?

Not exactly. There have been cases where tourists were denied entry and faced extended detention but they were held in U.S. facilities before being returned to their home countries.

There have been cases of incoming immigrants sent to foreign prisons without charges or trial. In March 2025 the U.S. administration flew 238 Venezuelan migrants from Texas directly to the "Terrorism Confinement Center" in El Salvador. The group included asylum seekers with active court cases, individuals who had entered the U.S. legally, and recent border crossers. The administration labeled them members of the "Tren de Aragua" gang to justify the move.

Not to be pedantic, but although the datacenters are running Nvidia hardware, Tesla did develop their own 20-core/3-npu high bandwidth chip for their cars. It's nowhere near the computational ability of any datacenter GPU, but at 150+ TOPS it's no slouch either.

>"The data provided is NOT a single playable Minecraft world, but rather a highly compressed collection of several world downloads of 2b2t."

I feel like HN needs to have a small model that compares post title to the article content and assigns it an accuracy score.

Have you looked at the BPI-R4? It's a pricier option than the OpenWRT One, but it has excellent hw acceleration for networking tasks. I am 90% sure I recall someone reporting using it for a 2.5Gbps PPPoE connection and it handled it. It's also supported by the OpenMPTCProuter project if you need network aggregation support.

FreshTomato is still active, and they are doing x86 builds now. I'm running it on an aging Netgear R7000 and it has been stable.