technology

Wow, thanks for the stickies! Love all the activity in this thread. I love our coding comrades!

Benchmarks: https://artificialanalysis.ai/models/deepseek-v4-flash

Recommended by Marxist tiktok creator @leessyndikaat, the article argues that Chinese AI technologies (described as a “Productivity Singularity") should be developed and disseminated globally as part of a revolutionary marxist strategy, building the material conditions for communism.

https://archive.ph/DMGaS

The battered hedge fund founded by former OpenAI researcher Leopold Aschenbrenner is unwinding many of its trades after big losses on artificial intelligence stocks and a bad bet against software stocks left it scrambling to raise cash, according to people familiar with the matter.

Goldman’s strategists reckon this means around one in 30 adults in the country, or 3.4% of the adult ⁠population margin called

Artificial intelligence labs are in a new arms race to buy up millions of rare books, slicing them open, scanning the pages and pulping the remains — sparking concerns that the last remaining copies of out-of-print texts are being destroyed on an industrial scale.

Kimi K3 is a massive open weights Mixture-of-Experts model with 2.8 trillion total parameters and 104 billion active parameters per token, and it is able to process text, images, and video natively within a 1 mil token context window. Combining Kimi Delta Attention for efficient long-sequence processing with Attention Residuals and Stable LatentMoE gave a 2.5 times improvement in scaling efficiency over their previous K2 model.

Kimi Linear which is a hybrid attention model designed to replace standard full attention while fixing the extreme memory and compute bottlenecks that happen during long context tasks and reinforcement learning. They built a custom linear attention module called Kimi Delta Attention and interleaved it with standard Multi-Head Latent Attention layers at a three to one ratio. It turns out that their custom layer improves on earlier gated linear models by using a channel-wise forget gate instead of a coarse head-wise one. The fine-grained gating gives the model much better control over its recurrent memory. They also came up with a chunkwise parallel algorithm that makes it incredibly hardware efficient compared to older matrix formulations.

https://archive.is/iozRD (bbc)

https://archive.ph/CRFok

I came across a project, “0day Rubbish”, that states it will continuously

worse still — we aren’t even talking about socialization. What Sanders is talking about is (partial) nationalization. The question we are debating is not “Should the workers take over?” but “Should the government?” It may have escaped the notice of those in that debate, but the current government is not particularly sympathetic to socialist interests. That Donald Trump and Bernie Sanders have both “called for the nationalization of AI firms” is presented by Guastella as a positive thing, as transcending the partisan divide. We think it represents a grave strategic error.

Interesting article about Medhurst and his use of GrapheneOS. Apparently he failed to turn off his phone before UK police boarded his plane and seized it from him, but the auto reboot feature protected his data and sources. The article also mentions that he’s now in the process of writing a massive guide to security and privacy for journalists.

linky

https://archive.ph/nqFsz

DDR5 pricing in Germany has been picking up pace again, according to the latest report by 3D Center. Unfortunately, this July has seen a 7% monthly price increase for DDR5 kits, on top of the already elevated prices, pushing DDR5 to new all-time highs. Despite stagnation between February and slight price decreases in March and June, July has continued the upward trend, now placing the average cost of DDR5 memory at a 448% increase over the prices recorded back in July 2025. In a single year, prices have more than quadrupled, mostly driven by massive price hikes between October 2025 and January 2026. PC enthusiasts have already faced significant affordability issues due to these price hikes, and there doesn’t seem to be an end to this trend.

Training LLMs requires thousands of accelerators that need to be wired together with ultra fast cables because they constantly share intermediate states and gradients during every single step of the optimization process. Building and maintaining a single compute cluster of that size is something that only large corps can afford to do.

Checking out a parallel copy of our Go repository and telling the AI to rewrite the whole thing in Zig while I work on something else just so I can keep my job. I hate this shit so much. My job has usage tracking and quotas. I don’t use it for actual work, I just spin it up and disregard the output. – An actual software engineer

Edit: I want to be clear because I’m not sure how much is jokes vs people not understanding. The robot teacher is not developed by the RealDoll team and does not, in fact, have a vagina.

https://www.scmp.com/tech/article/3361119/alibaba-says-newest-qwen-ai-model-second-only-anthropics-claude-fable-5

Full text: (Archive link: https://archive.ph/Sy4tV)

The researchers trained a trillion-parameter model using zero reinforcement learning to see how reasoning capabilities emerge at a massive scale without relying on human-annotated data. They found that pushing the parameter count to a trillion drastically improves both sample efficiency and the overall performance ceiling when compared to a smaller 104-billion parameter baseline strongly validating the concept that raw scale and computation eventually outpace hand-crafted human heuristics.

After many rumours about what OpenAI’s first hardware product would be—with rumours ranging from a speaker to an iPhone killer—the company has now revealed the Codex Micro, which simply put is a keypad. This is obviously not what anyone was expecting and at US$230 for what is largely a compact keypad from specialty keyboard maker Work Louder, it’s hard to see how this is going to be a big seller.