Artificial Intelligence (AI) research prides itself on openness, yet many so-called open source AI projects are not as “open” as they appear. The term open source traditionally implies full transparency and freedom to use, modify, and share software. In the context of AI, however, openness becomes complicated by proprietary datasets, model weights, and restrictive licenses. This disconnect has sparked growing concern among students, developers, and academics committed to ethical AI development and transparency in AI research. 

In this post, we explore why “open source” AI isn’t always truly open, what leading researchers and case studies reveal about this trend, and how the community is responding. We’ll also examine real-world academic case studies and provide a roundup of scholarly resources on the topic.

What Does “Open Source AI” Really Mean?

In software, open source means releasing source code under licenses that guarantee users the freedom to use, study, modify, and distribute it. For AI systems, defining “open source” is more complex. AI models consist of not just code, but also training data, learned parameters (model weights), and sometimes hardware dependencies. The Open Source Initiative (OSI), which maintains the canonical Open Source Definition for software, recently led a two-year effort to craft a definition of open source AI that extends these principles to AI’s unique components​. 

According to OSI, a truly open AI system must provide all of the following:

• Training data transparency: Access to details about the data used to train the AI, so others can understand what the model learned from and attempt to re-create it under the same conditions​.
• Open source code: The complete code (architectures, preprocessing, etc.) used to build and run the AI, released under an OSI-approved license​.
• Model weights and parameters: Access to the model’s learned weights and training settings, enabling others to use or fine-tune the exact model and verify its results​.

In essence, open source AI should uphold the “four freedoms” familiar from software: the freedom to use, examine, modify, and share the technology without undue restrictions. Yet many AI projects labeled “open” do not meet these criteria. For example, the OSI notes that Meta’s LLaMA 2, widely touted as an open-source AI model, fails to qualify because it withholds the training dataset and imposes usage restrictions that violate open license norms. Instead of unconditional transparency, Meta offered “resources available to some users under some conditions,” which OSI argues is fundamentally different from true open source. Without a universally accepted definition, powerful companies can exploit ambiguity in the term “open source AI,” potentially using it as marketing while limiting access in practice. This lack of clarity is precisely why OSI and academic stakeholders are working to reclaim the term and set stricter standards.

“Open source AI” should mean more than just downloadable code or model files – it implies complete transparency and freedom to build upon the work. Unfortunately, as we’ll see next, some organizations use the label without delivering the openness that researchers expect. This gap between the ideal and reality has given rise to the phenomenon of openwashing in AI.

When “Open” Isn’t Truly Open: Openwashing in AI

As AI booms, several high-profile companies have begun openwashing – claiming to be “open source” to gain goodwill or regulatory advantages while keeping critical parts of their AI systems closed. Openwashing is a play on “whitewashing,” referring to misleading use of the open-source label. Researchers warn that this trend is on the rise: one 2024 study surveying 45 generative AI systems found many were “open in name only,” providing perhaps model weights but withholding training data or key documentation (Liesenfeld & Dingemanse, 2024). Such models are often described as “open-weight” at best – the neural network parameters may be downloadable, but without the original data or full code, the model is not independently reproducible or fully modifiable​.

Why would companies do this? One motive is to sidestep impending regulations. The European Union’s AI Act (expected 2024) plans to exempt open source AI from certain strict requirements. By slapping an “open” label on a model (without actually meeting open criteria), providers hope to avoid oversight while still controlling their “secret sauce”. Stefano Maffulli, OSI’s director, calls this “openwashing” and notes it has the effect of compromising professional standards, moving AI development away from the core values of transparency and freedom that true open source embodies. Media excitement around releases like LLaMA 2 amplified the issue: headlines praised Meta for openness, often uncritically, even though the model lacked key open elements​. This PR strategy can mislead the public and policymakers into thinking a model is as inspectable as, say, Linux, when it’s not.

Openwashing’s consequences are substantial:

• Stifling Innovation: If big tech firms reap the reputational benefits of “open” without actually sharing knowledge, it undermines genuine open projects. “Large corporations can derive benefits from the trappings of open source without doing the requisite work,” effectively discouraging true openness. This disincentivizes the collaborative spirit that drives innovation​.
• Hurting Researchers: Scholars may assume a model labeled “open” is fully accessible, only to discover they cannot actually tinker with its internals or data. As one analysis notes, openwashing means researchers “can no longer count on being able to…study models even if they are advertised as open source” . This false sense of openness can waste academic effort and hinder reproducible science (more on this in the next section).
• Eroding Trust: Overuse of the term “open” for partially closed AI erodes the public’s understanding of AI. It blurs the line between truly transparent, ethical AI practices and corporate-controlled releases, potentially leading to complacency about accountability in AI development. An “open” label should signal trust and community oversight; openwashing turns it into a marketing buzzword, weakening its value.

Openwashing in AI is a concerning trend that both researchers and open-source advocates are pushing back against. Calls are growing for clearer standards and truth-in-labeling so that “open source AI” actually delivers on its promise. Next, we examine one root cause of this ambiguity – the complex maze of AI licensing and components – to understand how AI projects end up only partially open.

The Nuances of AI Licensing and Intellectual Property

Traditional open-source software licenses (like MIT or Apache) were designed for code, but AI systems comprise multiple components – source code, model weights, training datasets, and even pre-trained models – each of which can have separate legal treatment. This leads to nuanced AI licensing arrangements that can confuse even experienced developers. A model might have open-source code but proprietary data, or openly released weights under a restrictive usage license. For instance, Meta’s original LLaMA (2023) release provided model weights to selected researchers, but only under a non-commercial research agreement; the broader public had no legal access. LLaMA 2 (2023) went a step further by making weights downloadable to all, yet its license still restricts commercial use in certain cases (e.g. if the application has over 700 million users) and is tied to an Acceptable Use Policy​. The OSI has pointed out that such conditions violate the Open Source Definition’s criteria of no field-of-use or user discrimination – thus, “Meta’s LLaMA 2 license is not Open Source” by OSI standards. In other words, calling LLaMA 2 “open source” conflates availability with freedom. The model is accessible, but not with “everyone can do anything” freedom that true open licenses grant​.

Key aspects that complicate open licensing for AI include:

• Model Weights vs. Code: Some projects share code but not model weights—or vice versa—limiting reproducibility. Without both, retraining or modifying the model is nearly impossible.
• Data Transparency and Licensing: Training datasets are often undisclosed due to copyright or legal risks. Without data, researchers can’t verify results or audit bias, even if code and weights are available.
• Restrictive “Open” Licenses: Licenses like “research-only” or “ethical open” impose limits on use or users. These violate true open-source principles and create confusion around what “open” really means.

The licensing landscape for AI is a patchwork of compromises. It reflects legitimate concerns – from intellectual property to misuse prevention – but at the cost of openness and clarity. Without standardized licenses that cover code, data, and models together, “open” AI will remain a spectrum of partial transparency. These nuances have real consequences for researchers and practitioners, especially when it comes to reproducibility and trust in AI science, as we explore next.

Transparency and Reproducibility in AI Research

In scientific research, reproducibility is paramount: independent researchers must be able to verify and build upon published results. However, the rise of complex, large-scale AI models has introduced new reproducibility challenges. When a study touts an impressive result from a proprietary model or a model trained on a secret dataset, peers cannot fully validate those findings. This issue has become so acute that some have warned of a “reproducibility crisis” in AI. A cornerstone of rigorous science is that methods and materials (in this case, code and data) are shared; without them, results may as well be irreproducible magic.

Consider a telling incident: In early 2023, OpenAI’s Codex (a language model for coding) was used in numerous academic projects (over 100 papers) for code generation experiments. Codex was only available via OpenAI’s API – researchers did not have the model itself. When OpenAI abruptly shut down access to Codex with less than a week’s notice, those papers became impossible to replicate. One day, researchers could query Codex; the next, it was gone, and with it the ability to verify any prior results that depended on it​. OpenAI offered a limited, case-by-case research access program for Codex after an outcry, but this gated approach falls far short of the open availability needed for true scientific reproducibility (Kapoor & Narayanan, 2023). The Codex case starkly illustrates how privately controlled AI models act as fragile “research infrastructure” – subject to removal or change at any time​.

Empirical studies are now confirming what many suspected: openness correlates strongly with reproducibility. In a 2024 replication study, Gundersen et al. attempted to reproduce 30 influential AI research results. They found that 50% of the studies could be at least partially reproduced. Crucially, the availability of code and data was a dominant factor: when both code and data were shared, 86% of papers were fully or partially reproducible, whereas having only the data (but not code) saw reproducibility drop to 33%. 

Well-documented data was especially important – some attempts failed because datasets were poorly described or unavailable – underscoring that sharing data is as vital as sharing code. Another analysis by Stanford researchers of 10 prominent large models found a disappointing trend: most developers gave extremely limited information about their models. Only 3 out of 10 disclosed at least some of their training data inputs, and only 2 revealed even the model’s size (number of parameters), while almost none provided details to fully reproduce training.

The implications are clear:

• Lack of Verification: Without access to models or data, scientists cannot verify claims about an AI system’s performance, capabilities, or biases. Results might not generalize, or could even be false, but the community would have no way to know. This undermines trust in academic AI models and results.
• Slower Progress: Each group might waste time re-implementing others’ methods or chasing elusive results that were artifacts of closed data. Openness allows researchers to stand on each other’s shoulders rather than start from scratch. For example, when models like BLOOM (an open 176B-language model) or OpenLLaMA are released openly, researchers worldwide can experiment and fine-tune them for new tasks immediately, accelerating innovation​.
• Unrealistic Benchmarks: Closed-source model results might set benchmarks that others struggle to match because they don’t have access to the same resources or secrets. This can create a skewed perception of progress. Open models and open research tools help ensure that scientific benchmarks are achievable and not reliant on proprietary advantage.

The transparency in AI research that true open source provides is not just an idealistic goal—it’s a practical necessity for reliable science. When AI models are fully open, researchers can reproduce and scrutinize them, leading to more robust findings and faster collective advancement. As we look at the broader picture, it becomes evident that openness (or the lack thereof) has societal and ethical ripple effects too, influencing who gets to participate in AI and who benefits from it. In the next section, we turn to the importance of accessible AI development for education, collaboration, and ethical progress.

Democratizing AI: Accessibility, Collaboration, and Ethics

One of the original promises of open source software was democratization – empowering anyone, anywhere to contribute and benefit. In the AI arena, where cutting-edge models often require massive resources, democratization is both more challenging and more crucial. Truly open AI can help level the playing field, enabling universities, smaller companies, and individuals (like students and independent developers) to participate in advances that would otherwise be locked behind corporate walls. Conversely, when AI is proprietary, it concentrates power and capabilities in the hands of a few big players, raising ethical concerns about equity, accountability, and AI for education.

Accessible AI for Education and Research

Open source AI is a boon for learning. Students and professors can freely study the internals of open models, use them in projects, and even contribute improvements. For example, when Hugging Face and hundreds of researchers released the BLOOM model openly, it not only provided a powerful multilingual model to experiment with, but also came with an ethical charter and documentation that serve as educational resources. In contrast, if a model like GPT-4 remains a black box, students can only learn from it by observation, not by understanding its construction. 

Open models and datasets enable hands-on experience, which is vital for the next generation of AI scientists and engineers. Moreover, open platforms (like TensorFlow, PyTorch, and scikit-learn – all open source libraries) have become fundamental AI research tools in academia; they show that openness in tooling lowers barriers to entry. Extending this openness to the models themselves is the next step to ensure AI for education is truly accessible to all.

Global Collaboration and Inclusion 

Open AI projects invite collaboration from around the world. They allow talent in regions that may not have giant AI labs to still partake in cutting-edge research. For instance, the BigScience project that produced BLOOM was a year-long collaboration involving over 1000 researchers from diverse backgrounds, united by a shared open mission. This model of collaboration contrasts with secretive corporate development and has led to innovations in how we approach multilingual data and responsible AI. Similarly, the Allen Institute’s OLMo project openly publishes its models, data, and training logs, enabling external researchers to audit and contribute to model improvements. 

Such efforts exemplify AI collaboration in the open: communities pooling expertise and computing resources to achieve together what few could do alone. The result is not just a more diverse set of contributors, but also AI systems that are scrutinized from many angles, potentially making them more robust and fair. Openness thus ties directly into ethical AI development – with more eyes on the process, there’s a better chance to catch biases or issues early and to ensure the technology serves a broad set of users.

Modularity and Multi-Agent Systems 

Open-source AI components can be recombined in novel ways to push the frontiers of AI application. In cutting-edge domains like multi-agent systems and ensemble agents, researchers often assemble multiple AI models, each specialized in a task, to work in concert. Openness makes this feasible – modules with known internals and open APIs can be plugged together. For example, an ensemble decision system might use an open NLP model for language understanding, an open vision model for image recognition, and an open planning agent, orchestrating them as a team to solve complex problems. Klover’s own approach with AGD™ (Artificial General Decision Making) leverages such ensembles of specialized agents to augment human decision-making. 

This kind of modular AI thrives in an open ecosystem, where agents can be audited, swapped, or improved by the community. If each agent were a proprietary black box, combining them would be cumbersome and risky (due to unknown failure modes or licensing conflicts). Thus, open source not only helps individual projects, but also enables ensemble AI systems that are greater than the sum of their parts. In an educational context, students can learn more by experimenting with these modular multi-agent setups, constructing complete solutions from open building blocks.

Ethical Accountability 

Finally, openness aligns with the drive for ethical AI. When models are transparent, it’s easier to identify biases, understand failures, and ensure accountability. Researchers can audit an open model’s training data for representation issues or test it for unfair behavior, leading to more ethical AI development. By contrast, with closed models, we often have to take the developer’s word on trustworthiness. Open-source AI also invites the broader community – including social scientists, policymakers, and affected users – to be involved in auditing and shaping technology. 

This inclusive approach is key to human-centered AI development. For example, if an education-focused AI (say, a tutoring system) is open source, educators and school administrators can inspect it and adapt it to their needs, ensuring it aligns with educational values and privacy requirements. Openness thus fosters a culture of shared responsibility in AI, rather than placing all power (and potential blame) in the hands of a few companies.

Real World Case Studies (Academic Perspectives)

To ground the discussion, let’s examine a few real-world cases from the academic and research community illustrating the spectrum from truly open AI to “open” projects with caveats:

BLOOM (2022) – Large Language Model by BigScience 

BLOOM is a 176-billion-parameter multilingual language model developed by an international collaboration of researchers. It was released under an open-access license with its model weights, code, and an extensive paper detailing its training data (roots in 46 natural languages and 13 programming languages). BLOOM’s development was inclusive (over 1000 contributors) and it has an ethical charter guiding its use. 

This case showed that academic collaborations can produce cutting-edge AI that is fully open and competitive with industry models. By open-sourcing BLOOM, the project enabled researchers to study large-scale model behavior, conduct reproducible experiments in many languages, and even fine-tune it for specific applications without needing to train a giant model from scratch. BLOOM exemplifies how openness can drive innovation and inclusion in AI research.

OLMo (2024) – Fully Open Language Model by Allen Institute for AI 

The Allen Institute’s OLMo project set out to create fully transparent large language models to push the boundaries of openness. In 2024 they released OLMo 2, a 32-billion-parameter model that outperforms certain closed models like GPT-3.5 on benchmarks, while providing complete transparency. The team released everything: model weights, training code, detailed logs, and even intermediate checkpoints from the training process. 

Crucially, they also made the massive training dataset public, alongside documentation. This academic effort proved that being open doesn’t have to mean being behind in performance. OLMo’s release allows any researcher to replicate the training run or adapt the model – a level of openness that goes beyond even many “open-source” AI claims. It’s a powerful real-world demonstration that academic labs can lead in openness, setting a high bar for industry.

OpenAI Codex & Reproducibility (2021–2023) 

On the flip side, the saga of OpenAI’s Codex (mentioned earlier) serves as a cautionary tale for academic reliance on closed AI. Codex, a model that translates natural language to code, was hailed as a breakthrough and used in numerous research papers for tasks like code generation and AI-assisted programming. However, because neither the model nor its training data were ever open, researchers could only access it via a proprietary API. When OpenAI discontinued Codex in 2023, it left academia in the lurch – ongoing experiments and even published results could no longer be verified or extended by others. One analysis by Princeton researchers noted that “hundreds of academic papers would no longer be reproducible” due to this change​. 

The Codex case underscores the risk of “closed” AI in research: even widely used models can vanish or change, whereas truly open models (or open-source clones) could be self-hosted and preserved for reproducibility. In response, academic communities are now more cautious, often preferring open models (like Open Assistant or code models such as SantaCoder) to avoid being dependent on a single provider’s goodwill. (Sources: Kapoor & Narayanan, 2023)

Meta’s LLaMA (2023) – Leaked Weights and Community Response 

Meta’s LLaMA, mentioned earlier, provides an interesting in-between case. Initially, Meta released LLaMA’s pretrained weights only to academic partners under a restrictive research license. However, the weights were leaked online in early 2023, after which they spread widely among researchers. This unintended “release” led to a flurry of innovation in the open community: multiple research groups created fine-tuned versions, and LLaMA-based models began appearing in academic workshops. In July 2023, perhaps recognizing the enthusiasm, 

Meta launched LLaMA 2 as a purportedly open model (with some license restrictions as discussed). While not fully open by OSI standards, LLaMA 2 at least allowed broad access to a powerful model. The academic community’s embrace of the leaked LLaMA demonstrated a pent-up demand for open high-quality models. It also showed that when top-tier models are not openly available, researchers might resort to gray-area solutions to obtain them – a situation that could be avoided if such models were truly open to begin with.

Academic Case Studies Every AI Researcher Should Know

For those interested in digging deeper, here’s a roundup of influential academic and industry research on open source AI, licensing, and transparency. These sources provide further evidence, case studies, and frameworks from experts seeking to make AI more open and accountable:

• Maffulli, S. (2025) – “‘Open source’ AI isn’t truly open — here’s how researchers can reclaim the term,”Nature 640, 9.
  • A perspective from the OSI’s Executive Director (Stefano Maffulli) published in Nature, explaining how many AI models misuse the open-source label. It calls for clear definitions and warns that without true openness (including sharing training data and code), the scientific community may be stuck with unverifiable, corporate-controlled models. Introduces the concept of openwashing in AI and references upcoming regulations exempting open software.
• Liesenfeld, A., & Dingemanse, M. (2024).“Rethinking open source generative AI: Open-washing and the EU AI Act,” in ACM FAccT 2024 Conference.
  • A peer-reviewed study that defines a framework for assessing AI “openness” across 14 dimensions. The authors survey dozens of generative AI systems and find that openness is often partial – e.g., “many models are ‘open weight’ at best”. They discuss open-washing in depth, noting it is used to evade scrutiny, and detail how this practice harms innovation and research by undermining trust​.
• Gundersen, O. E., et al. (2024).“The unreasonable effectiveness of open science in AI: A replication study,” arXiv preprint arXiv:2412.17859.
  • Reports the results of a systematic attempt to replicate 30 highly-cited AI research papers. The study quantitatively demonstrates the impact of open science practices: when code and data were shared, the majority of results could be replicated, versus much lower reproducibility when only partial resources were shared. It highlights specific barriers (like inaccessible data or undisclosed hyperparameters) that cause replication failures, reinforcing the argument that open-source practices (sharing code, data, models) markedly improve reproducibility in AI research​.
• Stephens, M., et al. (2024).“Unraveling open source AI,”California Management Review (June 17, 2024).
  • An academic management perspective that offers a definition and framework to understand the ambiguities in the term “open source AI.” It discusses different combinations of open components (software, hardware, data, knowledge) and points out that some projects claim to be open while only releasing certain aspects (e.g. just the model weights) and holding back others (Stephens et al., 2024). The article uses examples like the Allen Institute’s OLMo as a project aiming to fully “open up” the AI training process​.
• Gent, E. (2024).“The tech industry can’t agree on what open-source AI means. That’s a problem,”MIT Technology Review (March 25, 2024).
  • A technology journalism piece that captures the state of the debate in early 2024. It notes how companies like Meta and personalities like Elon Musk were loudly promoting “open-source AI” (Musk even taking legal action against OpenAI for not being open). However, there was no consensus on the definition, leading OSI and others to step in (Gent, 2024). Gent explains the risk that without a clear definition, big firms could co-opt “open-source” to serve their interests, possibly cementing their dominance while giving the illusion of openness. The article also describes OSI’s effort, convening experts to craft a formal definition, foreshadowing the OSI definition release later in 2024 that set criteria for data, code, and weights​.

Together, these case studies illustrate a growing academic consensus: true openness in AI requires transparency across data, code, and usage rights—not just marketing claims. As the research shows, partial access undermines reproducibility, trust, and collaborative progress. For the next generation of AI researchers, these resources provide a vital roadmap for building ethical, accountable, and genuinely open AI systems.

MovingTowards Truly Open and Ethical AI

For Klover, as a leader in human-centered AI, these findings reinforce our mission. We believe that AI should empower people – whether it’s students learning with AI, developers building new solutions, or communities using AI for social good. That empowerment only happens when AI is accessible and trustworthy. Klover’s vision of ethical AI and our trademark AGD™ (Artificial General Decision Making™) paradigm prioritize transparency and collaboration. Rather than relying on a single monolithic AI, AGD™ utilizes an ensemble of specialized multi-agent systems working together, each of which can be understood, audited, and improved. This modular approach aligns with open principles: in an ensemble of AI agents, components need clear interfaces and accountability, much like open-source modules. By championing openness, we make it feasible to compose these ensemble agents safely and effectively, leading to more robust and fair outcomes.

Our commitment to openness is also about trust. Users and stakeholders should be able to know why an AI decision was made – and that’s far more achievable when the models and data are not black boxes. It’s why Klover supports initiatives for transparent model reporting and contributes to open research efforts. The implications of embracing true openness in AI are far-reaching: we can accelerate scientific discovery, democratize AI education, and ensure that AI systems reflect a broad array of values and perspectives. These are exactly the outcomes that align with Klover’s human-centric ethos and our goal of leveraging AI for better decisions and a better world.

Works Cited

Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., … & Zheng, X. (2016). TensorFlow: A system for large-scale machine learning. In 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16) (pp. 265–283). 

Cigionline. (2023, July 27). Open AI, closed sources: A transparency index for foundation models. CIGI. 

Gent, E. (2024, March 25). The tech industry can’t agree on what open-source AI means. That’s a problem. MIT Technology Review. 

Gundersen, O. E., Cappelen, O., Mølnå, M., & Nilsen, N. G. (2024). The unreasonable effectiveness of open science in AI: A replication study. arXiv preprint arXiv:2412.17859. https://arxiv.org/abs/2412.17859

Heaven, W. D. (2023, March 6). Meta’s new AI model is leaking everywhere—but it’s not open source. MIT Technology Review. 

Kapoor, S., & Narayanan, A. (2023, March 22). OpenAI’s policies hinder reproducible research on language models. AI Snake Oil. 

Le Scao, T., Fan, A., Pavlick, E., Akiki, C., Alam, T., Wang, Y., … & Rush, A. M. (2022). BLOOM: A 176B-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100. https://arxiv.org/abs/2211.05100

Liesenfeld, A., & Dingemanse, M. (2024). Rethinking open source generative AI: Open-washing and the EU AI Act. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’24). 

Maffulli, S. (2023, July 20). Meta’s LLaMA license is not open source. Open Source Initiative. 

Maffulli, S. (2025, March 27). ‘Open source’ AI isn’t truly open — here’s how researchers can reclaim the term. Nature, 640, 9. 

Meta AI. (2023, July 18). Introducing LLaMA 2: Open foundation and fine-tuned chat models. Meta AI Blog. 

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., … & Chintala, S. (2019). PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32. 

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., … & Duchesnay, E. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830. 

Sid Sijbrandij. (2023, June 5). AI weights are not open “source”. OpenCore Ventures Blog. https://www.opencoreventures.com/blog/ai-weights-are-not-open-source

Stephens, M., Esposito, M., Awamleh, R., Tse, T., & Goh, D. (2024, June 17). Unraveling open source AI. California Management Review. Walsh, P., Soldaini, L., Groeneveld, D., Lo, K., & Ammar, W. (2025). 2 OLMo 2 Furious: The next generation of fully open language models. arXiv preprint arXiv:2501.00656.