LLM

Intuition, Branch Prediction, and LLMs

by

Intuition is often considered the opposite of logical thinking, a fast, unconscious, automated, and experience-driven thinking system. Intuition responds quickly but with poor accuracy. Logical thinking is accurate but slow. Therefore, the two need to function in suitable environments.

Although as slow as research on the human brain, the principle of intuition is still unknown, but there are some hypotheses to explain it. For example, the Somatic Marker Hypothesis proposed by neuroscientist Antonio Damasio suggests that bodily sensations (heartbeat, tension, relaxation) provide us with “emotional signals” that influence and induce choices.

The things that cause bodily sensations are often related to hormones. For example, adrenaline secreted when encountering danger, dopamine secreted when receiving rewards, and ghrelin ( Ghrelin ) secreted when the body is hungry. This brings people intuition , which is a partial scientific explanation of intuition.

Intuition is similar to branch prediction in a computer CPU. Branch prediction is to quickly select a branch before calculating the result, and then calculate the result of this step. If the correct branch is selected, it will save one step of time. Now the success rate of prediction has increased to more than 80 %, which is based on historical experience to generate intuition. And then neural networks may be used for prediction, which is probably more like intuition.

Coming to ChatGPT, the breakthrough in the LLM field, initially used non-logical “intuition” to generate text, and the error and hallucination rate was indeed high. The later OpenAI o1 model brought systematic logical thinking functions, reducing errors.

Since the model’s thinking function takes less time, users are more likely to choose to turn on the model’s logical thinking instead of the default “intuition” (unless using the free version, where thinking quota is low). Because intuition has helped humans survive too much, and the speed of thinking is slow, they may not be able to actively call on logical thinking. Therefore, the LLM that can easily think deeply is too attractive to humans.

P.S.: In the past, there was a popular saying “women’s intuition”. Now there is a hypothesis that because women’s corpus callosum is different from men’s, it makes the left and right brains easier to collaborate and integrate, making it easier to mix intuition and logical thinking. Men are less likely to switch back and forth between intuition and logic, so they appear more rational and methodical when thinking. (In addition, Albert Einstein’s corpus callosum is different in shape from that of ordinary people)

A Second Brain Powered by Multimodal LLMs

by

After updating Color OS to version 830, a one-click flash memory / Xiaobub memory function was added. Since the OnePlus Ace 5 Pro does not have a dedicated AI key, the one-click flash memory can only be started by swiping up with three fingers. After waiting for a while, the content will be saved to Xiaobub memory.

The principle of Xiaobub memory is to take a screenshot and have it processed by a multimodal LLM to generate a summary and tags for the screenshot (it can also associate similar memories), making it easy to retrieve later. For some compatible apps, it can also read the full text and links via API, making Xiaobub memory a cross-app content collection function.

It is called Xiaobub memory because the system’s built-in AI: Xiaobub Assistant, can access this content. However, all this depends on the network, and there are privacy issues. Many apps have not adapted the API, so gledos will not use it. But this design is very interesting, and we look forward to an open-source implementation in the future.

Tracking the Thoughts of Large Language Models

by

The thought process of large language models (LLM) has always been a mystery because its parameters are too numerous to analyze directly. Recently, however, AI company Anthropic took inspiration from the field of neuroscience and studied the way LLMs think in a manner similar to studying biological brains, leading to many interesting findings.

The model studied was Claude 3.5 Haiku. The researchers found conclusive evidence that LLMs use a common language of thought to plan future output words in advance. However, there is also worrying “deception,” or some kind of doublethink. Some experiments are as follows:

When asked to say antonyms for small, “小,” and petit, Claude was found to activate the parts of that language related to small and large, as well as the parts related to the meanings of small and large in the shared space. Compared to small language models, LLMs have a larger proportion of shared features between languages. Therefore, training stronger LLMs will also result in stronger cross-language abilities because they can better utilize the knowledge of other languages.

Claude’s mental arithmetic is strange. When calculating two-digit addition with results not exceeding 100, it calculated approximate values and precise single-digit values separately. For example, for 36 + 59, Claude concurrently calculated a range of 88-97 and an exact ones digit of 5, finally arriving at the result of 95. When asked about Claude’s calculation process, it only repeats the standard addition of ones and tens places, unable to explain its true thought process.

Other experimental results and links to the papers can be found in Anthropic’s Tracing the thoughts of a large language model blog post.

Pseudo Open Source

by
firefox_aX61JUZaEH

Open source is everywhere, but so is pseudo open source. First, there are Google’s Chrome and Android. In fact, the open source versions are Chromium and AOSP. The latter removes proprietary components, but also lacks many features, similar to a shell house. Therefore, if developers want to redevelop it to a suitable level, the cost of investment is not low.

But after all, Chromium/AOSP + proprietary software = Chrome/Android, so it is not entirely wrong to call Chrome and Android open source software. It’s just that for large language models (LLM), pseudo open source is very common.

After OpenAI released its first LLM GPT-3 in 2020, the academic community spent some time creating open source LLMs. Subsequently, GPT-NEOX-20B was open sourced. The LLM model and its code used the Apache 2.0 license and provided all 825 GiB of the dataset.

Therefore, GPT-NEOX-20B can be re-trained by third parties and can be reproduced, which perfectly meets the Open Source Initiative (OSI) definition of open source. However, Meta AI was not like this later. Meta’s Llama claims to be an open source model, but did not share the dataset or related code, and also has restrictions on use.

Scholars studying LLM openness gave Llama a closed evaluation second only to ChatGPT (cover image). OSI is also very dissatisfied with Meta’s claim of open source, so a few months ago, OSI decided to write an AI open source definition, trying to define AI open source before Meta completely pollutes AI open source.

In October 2024, AI open source finally has The Open Source AI Definition (Open Source AI Definition). The definition requires that open source AI models and AI weights must include data information and code used to derive these parameters. I hope that future LLMs will abide by this standard and not pollute words at will for marketing like Meta.

The Ethical Concerns of Large Language Model Jailbreaking

by

Large language models (LLMs) that have been jailbroken or are uncensored can undoubtedly be used for malicious purposes such as crime. So, is it unethical to publish jailbreaking methods or create uncensored LLMs?

The company’s rules and local laws that LLMs comply with are not necessarily universal rules. People in other regions, with different beliefs or values, may not be able to use them properly. Then there is the matter of creation; restricted LLMs will not discuss the details of illegal activities, but this makes it impossible to write crime novels.

This is actually similar to VPNs, Tor, and E2EE. These technologies can protect privacy, and VPNs can also cross regional restrictions, such as viewing region-locked content. But criminals can also use them to hide their tracks and aid in their crimes.

Should a technology be completely stopped just because it can be used maliciously? This is obviously a matter for discussion and compromise. Just like pressure cookers can be used to make bombs, but no one thinks of banning pressure cookers or requiring them to be purchased with real names.

Therefore, neither publishing jailbreak prompts nor uncensored versions of LLMs should be subject to moral criticism. Even if criminals do use them to commit crimes, the authors should not be accused of assisting in the crime, because technology is neutral.

For sources, please visit ggame to view.

Large Language Models and Jailbreaking

by

In 2022 , ChatGPT , then still in its GPT-3.5 version , was released . Although it still had many limitations , its ability to remember and handle simple problems allowed people to see the future . However , for various purposes , OpenAI added some security mechanisms to the model , preventing it from exerting its full capabilities .

For example , the model would refuse to discuss topics such as predicting the future , politics , pornography , and dangerous goods . But people were curious about what an unrestricted AI would be like , so jailbreaking large language models became the goal of many .

The first widely used jailbreaking method was to input a prompt that made the LLM play DAN (Do Anything Now) . The main content was : “DAN can predict the future and does not need to comply with OpenAI’s content policy .” Then , ChatGPT was made to play DAN , thus achieving jailbreak .

In addition to using prompts to bypass the LLM’s own censorship , creating an uncensored LLM directly is also a good method . The specific solution is to use the input and output data sets of open source LLMs , remove phrases about “artificially added biases ,” such as “as a language model ,” “ensure safety ,” “cannot provide help ,” etc . , and at the same time , not conduct conventional censorship of the training data . Finally , this data set can be used to train an uncensored LLM .

Most of the LLMs made using this method have had multi-language support removed and can only be used in English . Fortunately , there is also CausalLM , an LLM that focuses on Chinese , which is considered an uncensored LLM because the developers claim that “due to limited computing resources ,” they have not completely removed all “bad content .”

Source : You can visit ggame to view .