[3/31 - 4/4] This Week's AI News

The AI industry has been innovating at a remarkable pace in recent years, with various companies developing and offering their own large language models (LLMs). While the debate over whether OpenAI will truly commit to open-sourcing is now a standard trope among industry insiders, Anthropic's research into Mechanistic Interpretability is gaining attention as a major topic aimed at unraveling the inner workings of 'black box' AI. Additionally, although Apple has introduced AI technology through tools like the voice assistant Siri under the name 'Apple Intelligence,' some say it has not fully met public expectations. Meanwhile, Amazon, which operates large-scale infrastructure centered on AWS, has finally unveiled its new 'AI agent' strategy and is investing heavily in research and development.

In this article, based on these latest trends, we will focus on four companies—OpenAI, Anthropic, Apple, and Amazon—and explain professional content as clearly as possible. We will delve into the background and objectives of these developments while citing insights obtained from dialogues between multiple researchers and engineers.

1. Will OpenAI really become 'open'?

1-1. The traditional OpenAI and criticisms of being 'non-open'

'OpenAI has 'open' in its name, but it is actually 'closed',' has long been joked about in various places. Indeed, the training data and weights for the latest models, including GPT-4, are almost entirely private, making them 'black box' entities accessible only via API. On the other hand, over the past few months, the 'open model' field has been rapidly heating up as Meta and various startups have successively released open-source or open-weight models. Perhaps in response to this trend, OpenAI CEO Sam Altman has also hinted that 'there is a possibility of releasing open-weight models in the future,' which has become a major topic of discussion both inside and outside the industry.

1-2. Discussion: Is true open-sourcing possible?

On the program 'Mixture of Experts,' CTO Chris Hay analyzes that 'the emergence of open competitors like DeepSeek may have prompted OpenAI to act.' He also points out that 'it is difficult to release a GPT-4 level model in a completely open form.' Given the intentions of major investors and high training costs, the view is that 'open weights,' where relatively small models or older generation weights are released on a limited basis, is a realistic compromise.

Furthermore, AI advocate Ash Minhas states, 'OpenAI, having received significant investment, has a responsibility to its investors. There are high hurdles to opening up large-scale models.' On the other hand, the opinion was also shared that 'even if large-scale models are opened up, if they cannot be linked to user-friendly experiences or added value, business superiority cannot be maintained,' drawing attention to how essential OpenAI's 'open' strategy will be in the future.

2. Anthropic's research on Mechanistic Interpretability

2-1. Attempts to unravel the 'black box'

A long-standing challenge in AI research has been the problem of 'not knowing why a model makes such inferences.' As neural network layers have become more layered and vast, the chain of internal weights and activation functions has become a 'black box' that is difficult for humans to intuitively understand.

Anthropic is tackling this challenge head-on and has recently published two papers. They experimentally present a method to visualize 'which parts of a neural network influence text generation and inference processes, and how.' On the program, the approach of 'analyzing the internal chain-of-thought (inference process) like a kind of 'graph' and tracking the correspondence with the final output' was evaluated as interesting.

2-2. Polysemantic neurons and the 'planning' of models

According to Anthropic's research, the existence of 'polysemantic neurons,' where a single neuron simultaneously represents multiple different concepts, is suggested. Furthermore, it is said that behavior resembling 'planning' what to output next during text generation has also been observed.

Chris Hay touched on an experimental example where 'when the word 'rabbit' was made unusable, the model replaced it with 'habit' to generate a rhyming poem,' and stated, 'Inside the model, there appears to be something like a clear plan or intent.'

However, it is premature to take these results as 'the model is thinking like a human.' The important point is how far we can mechanically and systematically decipher the complex patterns occurring inside the model. In the future, this interpretability research is considered to become increasingly important for introducing AI into high-risk areas such as medicine and finance.

3. AI challenges facing Apple

3-1. Expectations and current status of 'Apple Intelligence'

Apple has adopted machine learning technology in various devices, starting with Siri. However, there is a view that due to robust hardware designs like the iPhone and HomePod and a strong commitment to privacy, they are forced to be cautious about AI model update cycles and the implementation of massive models. It was also pointed out on 'Mixture of Experts' that 'Apple always prioritizes user experience and has a culture of disliking releases in an incomplete state.' Since AI inherently exhibits probabilistic behavior, it may be a difficult situation for Apple, which emphasizes 100% stable operation.

3-2. Will Apple still not lose? On the other hand, when asked 'will iPhone users switch to other companies' devices because of 'advanced AI features,' that is not necessarily the case. It is said that Apple, which has deep trust in design and hardware quality, will release AI features at a stage where they can be confident that 'the level of perfection is high.' There are also strong voices saying, 'Even if there is no big movement immediately, in the long term, there will definitely come a time when Apple gets serious.' There is a good chance that they will suddenly announce a revolutionary AI feature that redefines the industry one day. 4. Amazon's AI agent strategy and 'Nova Agents'

4-1. What is Amazon's newly launched 'Nova Agents'?

The program concluded by spotlighting Amazon's new AI agent platform, 'Nova Agents.' With the vast data from logistics hubs powered by AWS cloud infrastructure and robotics as a backdrop, Amazon is finally entering the agent field in earnest.
Of particular note is Amazon's stance of positioning itself as if it has launched an 'AGI Research Institute,' developing its own models while simultaneously coordinating with existing large language models. Until now, they had stuck to the cloud vendor position of 'any model is fine as long as it runs on AWS,' but it seems the 'significance of having in-house models' is being re-evaluated.

4-2. Synergies with AWS and Ecosystem Building

Amazon has already established a system for low-cost, high-efficiency training and inference of large-scale models through Bedrock and proprietary chips (such as Trainium). AI architect Aaron Baughman commented, 'Amazon has vast data and use cases in e-commerce, AWS, and robotics. If they layer agent capabilities on top of that, they can gain an immediate advantage.'
Furthermore, AI advocate Ash Minhas predicts, 'They will likely provide an "ecosystem" where users can freely customize by opening up SDKs and toolkits to the public.' Unlike Apple's route of keeping everything in-house, creating an open market is highly likely to be the strategy that leverages Amazon's strengths.

How far will OpenAI's 'open' shift go, the future of mechanistic interpretability pioneered by Anthropic, Apple's quiet battle, and Amazon's new AI agent strategy—all are important topics for forecasting the industry's future.

• OpenAI: While considering investors and commercial aspects, it is highly likely they will proceed with open-weighting some models under pressure from the open-source community. However, it remains unclear whether all cutting-edge models will be released.

• Anthropic: Research into mechanistic interpretability, which unravels the 'inscrutability' of deep learning, is extremely important for advancing AI use in high-risk areas such as medicine and law. The stance of sharing such basic research can contribute significantly to the development of the industry.

• Apple: A perfectionist culture as a hardware company and high UX requirements make the introduction of AI features cautious. However, one cannot deny the possibility that they will leverage their solid brand power and assets to announce innovative AI features all at once at a certain point.

• Amazon: With a background of abundant resources such as AWS, e-commerce, and robotics, the foundation for deploying agent technology on a large scale is in place. The route of promoting an ecosystem through the development of in-house models and the public release of SDKs will likely become dominant.

AI technology is permeating society's infrastructure beyond the realm of software. From voice assistants and content creation via generative AI to robotics and the optimization of cloud infrastructure, and even the automation of corporate operations, its use will advance in every situation from now on. Among these, issues such as 'interpretability,' 'privacy,' and 'model quality control' will undoubtedly attract even greater attention.

Who will make the 'next big move'—OpenAI, Anthropic, Apple, or Amazon? In any case, the new technologies and research results pioneered by these companies will likely lead the future AI industry and have the potential to change the way society functions.

Related Articles