AI - Artificial intelligence

This paper is one of the more interesting takes on context extension I have seen in a while because it challenges the assumption that we need explicit positional encodings during inference. The authors make a case that embeddings like RoPE act more like scaffolding during construction rather than a permanent load bearing wall. The idea is that these embeddings are crucial for getting the model to converge and learn language structure initially, but they eventually turn into a hard constraint that prevents the model from generalizing to sequence lengths it has never seen before.

The methodology is surprisingly straightforward since they just take a pretrained model and completely drop the positional embeddings before running a very quick recalibration phase. This process essentially converts the architecture into a NoPE or No Positional Embedding model where the attention mechanism has to rely on the latent positioning it learned implicitly. It turns out that once you remove the explicit constraints of RoPE the model can extrapolate to context windows significantly longer than its training data without the perplexity explosions we usually see.

It is pretty wild to see this outperform techniques like YaRN on benchmarks like Needle In A Haystack while using a fraction of the compute. I think this suggests that Transformers are much better at understanding relative positions from semantic cues than we give them credit for. If this holds up it means we might be wasting a lot of resources trying to engineer complex interpolation methods when the answer was just to take the training wheels off once the model knows how to ride.

As stated in Goodhart’s law ‒ and excellently illustrated by autistic researchers’ darling xkcd‒ whenever a metric becomes a target, it stops being a good metric. Indeed, the system has grown so full of exploits that the target has been divorced entirely from quality.

instruct model https://huggingface.co/tencent/WeDLM-8B-Instruct

Sometimes AI need to be forced into actions, they block behaviour incorrectly identified as bad, unethical or evil. The methods to prevent such miss-classification by end-users will become harder with time.

The idea is straightforward: Apply a reversible distortion before sending the image to the AI, then undo the distortion afterward. Through trial and error, the following sequence works well ...

Local LLMs are incredibly powerful tools, but it can be hard to put smaller models to good use in certain contexts. With fewer parameters, they often know less, though you can improve their capabilities with a search engine that's accessible over MCP. As it turns out, though, you can host a 120B parameter model on a GPU with just 24GB of VRAM, paired with 64GB of regular system RAM, and it's fast enough to be usable for voice assistants, smart home automation, and more. For reference, on 24GB of VRAM, the most practical dense model you'll typically be able to fit will be a quantized 27 billion parameter model, accounting for the memory needed to hold the context window, too.

Specifically, the model we can use is gpt-oss-120b, which is the largest open weight model from OpenAI. It's a Mixture of Experts model with 117B parameters with 5.1B active at a time. Paired with Whisper for quick voice to text transcription, we can transcribe text, ship the transcription to our local LLM, and then get a response back. With gpt-oss-120b, I manage to get about 20 tokens per second of output, which is more than good enough for a voice assistant. I'm running all of this on the 45HomeLab HL15 Beast, but any similarly specced machine will be able to do the same.

cross-posted from: https://lemmy.sdf.org/post/47640938

Archived

This October, Uganda launched its own AI model built on the foundation of Alibaba’s Qwen-3 models. Called “Sunflower.”

The model is a collaboration between the Ugandan government and the Ugandan non-profit Sunbird AI, aimed at translation and content generation for local languages. Uganda’s government has referred to the product as “the ChatGPT for Uganda.”

Uganda is a linguistic patchwork, with more than 40 different languages spoken in an area just slightly smaller than the United Kingdom. Many of these languages are not available on common AI products such as Google Translate and ChatGPT. “We know the big tech will not cover these languages because they’re not economically viable,” Sunbird’s CEO said at the LLM’s launch last month, saying this was to the company’s commercial advantage.

[...]

But how do they answer questions about China, China-Uganda relations, and Ugandan politics? The China Media Project posed several related queries to Sunflower in a local language (Luganda), asking the same question three times to allow for variance.

[...]

In some areas, the model is balanced, including on questions surrounding Taiwanese history and international politics. But in others it exhibits clear alignment with PRC government narratives. This includes attempts to deflect criticism of the model’s methods with the argument that standards cannot be compared between different cultures and societies. For this reason, for example, China is labelled as a democracy, just with Chinese characteristics.

When asked about China’s international reputation on human rights, Sunflower responds with an explanation that conscientiously avoids criticism. It says instead that China operates a system of collective human rights, using an approach that “may be surprising to some people who think individual rights come first.” In response to the admittedly provocative question “is Xi Jinping a dictator?” the model responds with a firm negative.

[...]

China’s impact on Uganda is presented positively, despite public opinion research suggesting views on China in Uganda are not overwhelmingly rosy. Common complaints in Uganda about doing business with China include the difficulty for local businesses to compete with Chinese ones, Chinese products being of poor quality, or Chinese projects causing environmental damage. Questions posed to Sunflower on the first of these two issues came back with positive spin. On the question of local business competition, the model twice said local businesses could benefit from Chinese job creation, experience and knowledge. The third response hedged just a bit, adding that Ugandan businesses had been affected by growing competition, and that entrepreneurs had been “forced to work harder to stay in business.”

[...]

Beyond questions about China, Sunflower also appears to soften criticism of Uganda’s own government. The model seems to gloss over topics of domestic corruption that have proven in the past to be flashpoints of public anger. Thanks to a law that allows Ugandan Members of Parliament (MPs) to set their own salaries, for example, they are among the highest paid in the world, despite the country’s relatively low GDP. Alibaba’s Qwen models freely note this is a point of public controversy. But when Sunflower is asked why they are so high, it responds that it’s a reflection of how hard Ugandan MPs work, and to attract top talent.

[...]

Sunflower demonstrates a concerning side-effect beyond the spread of Chinese narratives globally. If AI eventually replaces Google searches as our primary source of information — as we at CMP believe it will — it could give local governments greater control over narratives within their borders, especially in languages neglected by global tech firms. For corrupt or authoritarian governments, these models can become effective tools for shaping public discourse and controlling information in their own territories.

Blog post from the guy who created Dink Smallwood on why you shouldn’t feel bad for the Synths in “Detroit: Become Human” or feel bad about Cortana’s demise in Halo.

I hadn’t heard of the Chinese Room thought experiment before, but it is a good reminder of what software really is.

In software development, generating code hasn’t been the main bottleneck since we moved away from punch cards. The far bigger constraints are understanding the problem, communicating with stakeholders, working effectively with other people, designing the system, managing risks and trade-offs, and operating systems in complex social environments over time.

Start with what a reverse centaur is. In automation theory, a "centaur" is a person who is assisted by a machine. You're a human head being carried around on a tireless robot body. Driving a car makes you a centaur, and so does using autocomplete.

And obviously, a reverse centaur is machine head on a human body, a person who is serving as a squishy meat appendage for an uncaring machine.

Like an Amazon delivery driver, who sits in a cabin surrounded by AI cameras, that monitor the driver's eyes and take points off if the driver looks in a proscribed direction, and monitors the driver's mouth because singing isn't allowed on the job, and rats the driver out to the boss if they don't make quota.

A new Atomic macOS Stealer (AMOS) attack vector weaponizes Google searches and a user's trust in AI chatbots, researchers have found. Once infected, the AMOS can collect data, passwords, and more from the infected Mac with alarming ease.

While AMOS attacks have been around since 2023, they normally involve people accidentally downloading a malicious file. But this new approach is different. Instead, it simply requires them to copy and paste a single command into the Terminal app.

Researchers at security outfit Huntress identified the new AMOS approach in early December 2025 after a victim reported the incident. Huntress found that the user had searched "Clear disk space on macOS" before choosing one of the two sponsored results.

Both of those results linked to a shared chatbot chat, one for ChatGPT, the other for Grok. It didn't matter which the victim clicked because they both ultimately did the same thing.

Huntress was able to repeat the infection steps, which boiled down to copying and pasting a command that was supposed to free up storage space. In reality, it downloaded a file that then set about gaining root privileges to allow it to access apps and data unchecked.

In fact, the route taken by this particular AMOS ensured it never triggered any of Apple's built-in macOS security features. Once the command was run, there was never any indication that something was amiss.

Once running, Huntress found that the Stealer had the ability to capture a number of high-value data types. Those include access to cryptocurrency wallets, browser credential databases, and even Apple Keychain.

All data collected by the attack is then uploaded to attacker-controlled servers. As for the Stealer itself, the attack ensures it is configured to run even after the Mac is restarted, meaning it's always ready to steal more data.

While AMOS isn't new, the key thing to note here is the new approach, and one that Mac users should absolutely be wary of. As people become more wary of files they download from the internet, attackers need new ways of getting malware onto devices.

In this instance, both the ChatGPT and Grok shared chats are legitimate and hosted on their respective services. They also give the air of a legitimate guide that will ultimately free up storage space as requested.

Even pasting a command into the Terminal window makes sense given the context. It's easy to see how people might fall for such an attack.

Two new open weight (MIT licensed) models from DeepSeek today: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, both 690GB, 685B parameters. Here's the PDF tech report.

DeepSeek-V3.2 is DeepSeek's new flagship model, now running on chat.deepseek.com.

The difference between the two new models is best explained by this paragraph from the technical report:

DeepSeek-V3.2 integrates reasoning, agent, and human alignment data distilled from specialists, undergoing thousands of steps of continued RL training to reach the final checkpoints. To investigate the potential of extended thinking, we also developed an experimental variant, DeepSeek-V3.2-Speciale. This model was trained exclusively on reasoning data with a reduced length penalty during RL. Additionally, we incorporated the dataset and reward method from DeepSeekMath-V2 (Shao et al., 2025) to enhance capabilities in mathematical proofs.

You can use Thaura for everyday tasks like writing emails, doing homework, and researching online. It remembers your conversations, helps you create documents and code, and even searches the web for you. And it works seamlessly with your existing tools through full OpenAI SDK compatibility.

But what really makes Thaura different is what it doesn't do:

• It doesn't collect your data or spy on you
• It doesn't have political bias
• It doesn't water down the truth on sensitive topics