-一位谷歌内部研究员泄露了一份文件，提出了一些讨论谷歌员工认为值得关注的问题
-文件表示谷歌在领先技术研究方面正在面临挑战，因为开源社区正在快速发展
-开源社区的语言模型（LLMs）解决了过往研究无法解决的问题，同时解决了模型更新中的缩放问题
-LLMs的进步将直接影响到谷歌的商业策略，因为用户可以找到更好的免费模型
-谷歌因为持有商业机密，保持技术竞争优势变得更加困难
-开源模型的创新正在发挥作用，谷歌应该学习并与开源社区合作

原文地址：

https://www.semianalysis.com/p/google-we-have-no-moat-and-neither

我们没有护城河We Have No Moat

OpenAI 也没有And neither does OpenAI

我们在 OpenAI 做了很多仔细的研究。谁将跨越下一个里程碑？下一步将是什么？We’ve done a lot of looking over our shoulders at OpenAI. Who will cross the next milestone? What will the next move be?

但令人不安的事实是，But the uncomfortable truth is, 我们没有能力赢得这场军备竞赛，OpenAI 也没有we aren’t positioned to win this arms race and neither is OpenAI。在我们争吵不休的时候，第三派一直在悄悄地吃我们的午餐。. While we’ve been squabbling, a third faction has been quietly eating our lunch.

当然，我说的是开源。说白了，他们在舔我们。I’m talking, of course, about open source. Plainly put, they are lapping us. 我们认为是“主要开放问题”的东西今天已经解决并掌握在人们手中。Things we consider “major open problems” are solved and in people’s hands today. 仅举几个：Just to name a few:

• LLMs on a Phone：LLMs on a Phone: 人们在 Pixel 6 上以People are running foundation models on a Pixel 6每秒 5 个代币的速度运行基础模型。 at 5 tokens / sec.
• 可扩展的个人 AI：Scalable Personal AI: 您可以在晚上在笔记本电脑上微调个性化 AI。You can finetune a personalized AI on your laptop in an evening.
• 负责任的发布：Responsible Release: 与其说是“解决”，不如说是“避免”。This one isn’t “solved” so much as “obviated”. 整个网站充满了艺术模型，没有任何限制There are entire websites full of art models with no restrictions whatsoever，文字也, and text is 不甘落后。not far behind.
• 多模态：Multimodality: 当前的多模态 ScienceQA SOTA 训练时间为一小时The current multimodal ScienceQA SOTA was trained in an hour。.

虽然我们的模型在质量方面仍然略有优势，但差距While our models still hold a slight edge in terms of quality, the 正在以惊人的速度缩小gap is closing astonishingly quickly。开源模型更快、更可定制、更私密且功能更强大。他们. Open-source models are faster, more customizable, more private, and pound-for-pound more capable. They are 用 100 美元和 13B 的参数做事，doing things with $100 and 13B params而我们在 1000 万美元和 540B 的参数下苦苦挣扎。他们在几周内完成，而不是几个月。这对我们有深远的影响： that we struggle with at $10M and 540B. And they are doing so in weeks, not months. This has profound implications for us:

• 我们没有秘方。We have no secret sauce. 我们最大的希望是学习 Google 以外其他人正在做的事情并与之合作。我们应该优先考虑启用 3P 集成。Our best hope is to learn from and collaborate with what others are doing outside Google. We should prioritize enabling 3P integrations.
• 当免费的、不受限制的替代品在质量上具有可比性时，人们不会为受限制的模型付费。People will not pay for a restricted model when free, unrestricted alternatives are comparable in quality. 我们应该考虑我们真正的增值在哪里。We should consider where our value add really is.
• 巨型模型正在减慢我们的速度。Giant models are slowing us down. 从长远来看，最好的模型是那些In the long run, the best models are the ones

  可以快速迭代。既然我们知道在 <20B 参数范围内有什么可能，我们应该做一些小的变体而不是事后才想到。which can be iterated upon quickly. We should make small variants more than an afterthought, now that we know what is possible in the <20B parameter regime.

发生了什么What Happened

At the beginning of March the open source community 3 月初，随着 Meta 的 LLaMA 被泄露给公众，开源社区得到了他们第一个真正有能力的基础模型。got their hands on它没有说明或对话调整，也没有 RLHF。尽管如此，社区立即理解了他们所获得的东西的重要性。 their first really capable foundation model, as Meta’s LLaMA was leaked to the public. It had no instruction or conversation tuning, and no RLHF. Nonetheless, the community immediately understood the significance of what they had been given.

随之而来的是巨大的创新涌现，主要发展之间的间隔只有几天（完整细目请参见时间表）。我们在这里，仅仅一个月后，有A tremendous outpouring of innovation followed, with just days between major developments (see The Timeline for the full breakdown). Here we are, barely a month later, and there are variants with 指令调整instruction tuning、, 量化quantization、, 质量改进quality improvements、, 人工评估human evals、, 多模态multimodality、, RLHFRLHF等的变体，其中许多是相互建立的。, etc. etc. many of which build on each other.

最重要的是，Most importantly, 他们已经解决了任何人都可以修补的扩展问题they have solved the scaling problem。许多新想法都来自普通人。培训和实验的门槛已经从一个主要研究机构的总产出下降到一个人、一个晚上和一台功能强大的笔记本电脑。 to the extent that anyone can tinker. Many of the new ideas are from ordinary people. The barrier to entry for training and experimentation has dropped from the total output of a major research organization to one person, an evening, and a beefy laptop.

为什么我们可以预见它的到来Why We Could Have Seen It Coming

在许多方面，这对任何人来说都不足为奇。开源 LLM 的当前复兴紧随图像生成的复兴之后。相似之处并没有在社区中消失，许多人称这是法学硕士的“In many ways, this shouldn’t be a surprise to anyone. The current renaissance in open source LLMs comes hot on the heels of a renaissance in image generation. The similarities are not lost on the community, with many calling this the “稳定扩散时刻Stable Diffusion moment”。” for LLMs.

在这两种情况下，低成本的公众参与都是通过一种成本低得多的微调机制实现的，称为In both cases, low-cost public involvement was enabled by a vastly cheaper mechanism for fine tuning called 低秩适应low rank adaptation或 LoRA，并结合了规模上的重大突破（, or LoRA, combined with a significant breakthrough in scale (图像合成的潜在扩散，LLM 的latent diffusionChinchilla ）。在这两种情况下，获得足够高质量的模型都引发了世界各地个人和机构的一系列想法和迭代。在这两种情况下，这都很快超过了大型企业。 for image synthesis, Chinchilla for LLMs). In both cases, access to a sufficiently high-quality model kicked off a flurry of ideas and iteration from individuals and institutions around the world. In both cases, this quickly outpaced the large players.

这些贡献在图像生成领域至关重要，使 Stable Diffusion 走上了与 Dall-E 不同的道路。拥有一个开放的模型导致了These contributions were pivotal in the image generation space, setting Stable Diffusion on a different path from Dall-E. Having an open model led to 产品集成product integrations、, 市场marketplaces、, 用户界面user interfaces和, and Dall-E 没有发生的创新。innovations that didn’t happen for Dall-E.

效果是显而易见的：The effect was palpable: 与 OpenAI 解决方案相比，在文化影响方面迅速占据主导地位，后者变得越来越无关紧要。rapid domination法学硕士是否会发生同样的事情还有待观察，但广泛的结构要素是相同的。 in terms of cultural impact vs the OpenAI solution, which became increasingly irrelevant. Whether the same thing will happen for LLMs remains to be seen, but the broad structural elements are the same.

我们错过了什么What We Missed

推动开源最近取得成功的创新直接解决了我们仍在努力解决的问题。更多地关注他们的工作可以帮助我们避免重新发明轮子。The innovations that powered open source’s recent successes directly solve problems we’re still struggling with. Paying more attention to their work could help us to avoid reinventing the wheel.

LoRA 是一种非常强大的技术，我们可能应该更加关注它LoRA is an incredibly powerful technique we should probably be paying more attention to

LoRALoRA 通过将模型更新表示为低秩分解来工作，这将更新矩阵的大小减少了多达数千倍。这允许以一小部分成本和时间进行模型微调。能够在几个小时内在消费类硬件上个性化语言模型是一件大事，works by representing model updates as low-rank factorizations, which reduces the size of the update matrices by a factor of up to several thousand. This allows model fine-tuning at a fraction of the cost and time. Being able to personalize a language model in a few hours on consumer hardware is a big deal, 特别particularly是对于 for 涉及近实时整合新的和多样化的知识的愿望aspirations that involve incorporating new and diverse knowledge in near real-time。事实上，这项技术的存在在谷歌内部并未得到充分利用，尽管它直接影响了我们一些最雄心勃勃的项目。. The fact that this technology exists is underexploited inside Google, even though it directly impacts some of our most ambitious projects.

从头开始重新训练模型是一条艰难的道路Retraining models from scratch is the hard path

LoRA 如此有效的部分原因在于——与其他形式的微调一样——它是可堆叠的。可以应用指令调整等改进，然后在其他贡献者添加对话、推理或工具使用时加以利用。虽然单独的微调是低秩的，但它们的总和不需要，允许模型的全秩更新随着时间的推移而累积。Part of what makes LoRA so effective is that – like other forms of fine-tuning – it’s stackable. Improvements like instruction tuning can be applied and then leveraged as other contributors add on dialogue, or reasoning, or tool use. While the individual fine tunings are low rank, their sum need not be, allowing full-rank updates to the model to accumulate over time.

这意味着随着新的更好的数据集和任务的出现，模型可以以低廉的成本保持最新，而无需支付全面运行的成本。This means that as new and better datasets and tasks become available, the model can be cheaply kept up to date, without ever having to pay the cost of a full run.

相比之下，从头开始训练巨型模型不仅会丢弃预训练，还会丢弃在顶部进行的任何迭代改进。在开源世界中，这些改进很快就会占据主导地位，这使得全面重新培训的成本极其高昂。By contrast, training giant models from scratch not only throws away the pretraining, but also any iterative improvements that have been made on top. In the open source world, it doesn’t take long before these improvements dominate, making a full retrain extremely costly.

我们应该考虑每个新应用程序或想法是否真的需要一个全新的模型。如果我们真的有重大的架构改进来阻止直接重用模型权重，那么我们应该投资于更积极的蒸馏形式，使我们能够尽可能多地保留上一代的功能。We should be thoughtful about whether each new application or idea really needs a whole new model. If we really do have major architectural improvements that preclude directly reusing model weights, then we should invest in more aggressive forms of distillation that allow us to retain as much of the previous generation’s capabilities as possible.

如果我们可以在小型模型上更快地迭代，那么从长远来看，大型模型的能力并不会更强Large models aren’t more capable in the long run if we can iterate faster on small models

对于最流行的模型尺寸，LoRA 更新的生产成本非常低（约 100 美元）。这意味着几乎任何有想法的人都可以产生一个想法并进行传播。一天之内的培训时间是常态。以这样的速度，所有这些微调的累积效应很快就会克服尺寸劣势。事实上，就工程师工时而言，这些模型的改进速度大大超过了我们使用最大变体所能做的，而且最好的已经在很大程度上LoRA updates are very cheap to produce (~$100) for the most popular model sizes. This means that almost anyone with an idea can generate one and distribute it. Training times under a day are the norm. At that pace, it doesn’t take long before the cumulative effect of all of these fine-tunings overcomes starting off at a size disadvantage. Indeed, in terms of engineer-hours, the pace of improvement from these models vastly outstrips what we can do with our largest variants, and the best 与 ChatGPT 没有区别are already largely indistinguishable from ChatGPT。. 专注于维护地球上一些最大的模型实际上使我们处于劣势。Focusing on maintaining some of the largest models on the planet actually puts us at a disadvantage.

数据质量比数据大小更好Data quality scales better than data size

其中许多项目都通过Many of these projects are saving time by 在小型、高度精选的数据集上进行训练来节省时间training on small, highly curated datasets。这表明数据缩放法则具有一定的灵活性。. This suggests there is some flexibility in data scaling laws. The existence of such datasets follows from the line of thinking in 此类数据集的存在遵循了Data Doesn’t Do What You ThinkData Doesn’t Do What You Think中的思路，并且它们正迅速成为在 Google 之外进行培训的标准方式。这些数据集是使用合成方法（例如从现有模型中过滤最佳响应）和从其他项目中提取的，这两种方法在谷歌都不占主导地位。, and they are rapidly becoming the standard way to do training outside Google. These datasets are built using synthetic methods (e.g. filtering the best responses from an existing model) and scavenging from other projects, neither of which is dominant at Google. 幸运的是，这些高质量的数据集是开源的，因此可以免费使用。Fortunately, these high quality datasets are open source, so they are free to use.

直接与开源竞争是一个失败的主张Directly Competing With Open Source Is a Losing Proposition

最近的进展对我们的业务战略有直接、直接的影响。This recent progress has direct, immediate implications for our business strategy. 如果没有免费、高质量的替代品，谁会为有使用限制的 Google 产品付费？Who would pay for a Google product with usage restrictions if there is a free, high quality alternative without them?

我们不应该期望能够赶上。现代And we should not expect to be able to catch up. The 互联网在开源上运行modern internet runs on open source是有原因的。开源有一些我们无法复制的显着优势。 for a reason. Open source has some significant advantages that we cannot replicate.

我们需要他们胜过他们需要我们We need them more than they need us

保守我们的技术秘密始终是一个脆弱的提议。谷歌研究人员正定期前往其他公司，因此我们可以假设他们知道我们所知道的一切，并且只要该渠道开放，他们就会继续这样做。Keeping our technology secret was always a tenuous proposition. Google researchers are leaving for other companies on a regular cadence, so we can assume they know everything we know, and will continue to for as long as that pipeline is open.

但是，由于法学硕士的尖端研究是负担得起的，因此保持技术竞争优势变得更加困难。世界各地的研究机构都在相互借鉴，以广度优先的方式探索解决方案空间，这远远超出了我们自身的能力。我们可以在外部创新冲淡其价值时努力保守秘密，或者我们可以尝试相互学习。But holding on to a competitive advantage in technology becomes even harder now that cutting edge research in LLMs is affordable. Research institutions all over the world are building on each other’s work, exploring the solution space in a breadth-first way that far outstrips our own capacity. We can try to hold tightly to our secrets while outside innovation dilutes their value, or we can try to learn from each other.

个人不受与公司相同程度的许可证约束Individuals are not constrained by licenses to the same degree as corporations

这种创新的大部分发生在 Meta 泄露的模型权重之上。Much of this innovation is happening on top of the leaked model weights from Meta. While this will inevitably change as 虽然随着真正开放的模型变得更好，truly open models这将不可避免地发生变化，但关键是他们不必等待。“个人使用”提供的法律保护和起诉个人的不切实际意味着个人可以在这些技术炙手可热的时候使用它们。 get better, the point is that they don’t have to wait. The legal cover afforded by “personal use” and the impracticality of prosecuting individuals means that individuals are getting access to these technologies while they are hot.

成为您自己的客户意味着您了解用例Being your own customer means you understand the use case

浏览人们在图像生成空间中创建的模型，从动漫生成器到 HDR 风景，创意源源不断。这些模型由深深沉浸在其特定子流派中的人们使用和创建，赋予我们无法企及的知识深度和同理心。Browsing through the models that people are creating in the image generation space, there is a vast outpouring of creativity, from anime generators to HDR landscapes. These models are used and created by people who are deeply immersed in their particular subgenre, lending a depth of knowledge and empathy we cannot hope to match.

拥有生态系统：让开源为我们工作Owning the Ecosystem: Letting Open Source Work for Us

矛盾的是，所有这一切的一个明显赢家是 Meta。因为泄露的模型是他们的，所以他们有效地获得了整个星球的免费劳动力。由于大多数开源创新都发生在他们的架构之上，因此没有什么能阻止他们将其直接整合到他们的产品中。Paradoxically, the one clear winner in all of this is Meta. Because the leaked model was theirs, they have effectively garnered an entire planet’s worth of free labor. Since most open source innovation is happening on top of their architecture, there is nothing stopping them from directly incorporating it into their products.

拥有生态系统的价值怎么强调都不为过。The value of owning the ecosystem cannot be overstated.谷歌本身已在其开源产品（如 Chrome 和 Android）中成功使用了这种范例。通过拥有发生创新的平台，谷歌巩固了自己作为思想领袖和方向制定者的地位，获得了塑造比自身更宏大的想法的能力。 Google itself has successfully used this paradigm in its open source offerings, like Chrome and Android. By owning the platform where innovation happens, Google cements itself as a thought leader and direction-setter, earning the ability to shape the narrative on ideas that are larger than itself.

我们对模型的控制越严密，我们做出开放替代方案的吸引力就越大。The more tightly we control our models, the more attractive we make open alternatives. 谷歌和 OpenAI 都倾向于防御性地倾向于发布模式，使他们能够严格控制其模型的使用方式。但这种控制是虚构的。任何想将 LLM 用于未经批准的目的的人都可以简单地选择免费提供的模型。Google and OpenAI have both gravitated defensively toward release patterns that allow them to retain tight control over how their models are used. But this control is a fiction. Anyone seeking to use LLMs for unsanctioned purposes can simply take their pick of the freely available models.

谷歌应该让自己成为开源社区的领导者，通过与更广泛的对话合作而不是忽视来带头。这可能意味着采取一些不舒服的步骤，比如发布小型 ULM 变体的模型权重。这必然意味着放弃对我们模型的一些控制。但这种妥协是不可避免的。我们不能希望既推动创新又控制创新。Google should establish itself a leader in the open source community, taking the lead by cooperating with, rather than ignoring, the broader conversation. This probably means taking some uncomfortable steps, like publishing the model weights for small ULM variants. This necessarily means relinquishing some control over our models. But this compromise is inevitable. We cannot hope to both drive innovation and control it.

结语：OpenAI 怎么样？Epilogue: What about OpenAI?

考虑到 OpenAI 当前的封闭政策，所有这些关于开源的讨论都会让人觉得不公平。如果他们不愿意，我们为什么要分享？但事实是，我们已经以源源不断的挖角高级研究人员的形式与他们分享一切。在我们阻止这种趋势之前，保密是一个有争议的问题。All this talk of open source can feel unfair given OpenAI’s current closed policy. Why do we have to share, if they won’t? But the fact of the matter is, we are already sharing everything with them in the form of the steady flow of poached senior researchers. Until we stem that tide, secrecy is a moot point.

最后，And in the end, OpenAI 并不重要OpenAI doesn’t matter。他们在相对于开源的态度上犯了与我们相同的错误，他们保持优势的能力必然受到质疑。除非他们改变立场，否则开源替代品可以并且最终会使它们黯然失色。至少在这方面，我们可以迈出第一步。. They are making the same mistakes we are in their posture relative to open source, and their ability to maintain an edge is necessarily in question. Open source alternatives can and will eventually eclipse them unless they change their stance. In this respect, at least, we can make the first move.

时间线The Timeline

2023 年 2 月 24 日 – LLaMA 启动Feb 24, 2023 – LLaMA is Launched

Meta 推出 LLaMA Meta launches LLaMA，开源代码，但不开源权重。在这一点上，LLaMA 没有调整指令或对话。与当前的许多模型一样，它是一个相对较小的模型（可用于 7B、13B、33B 和 65B 参数），已经训练了相对较长的时间，因此相对于它的大小来说相当有能力。, open sourcing the code, but not the weights. At this point, LLaMA is not instruction or conversation tuned. Like many current models, it is a relatively small model (available at 7B, 13B, 33B, and 65B parameters) that has been trained for a relatively large amount of time, and is therefore quite capable relative to its size.

2023 年 3 月 3 日 – 不可避免的事情发生了March 3, 2023 – The Inevitable Happens

一周之内，Within a week, LLaMA 就被泄露给了公众LLaMA is leaked to the public。对社区的影响不容小觑。现有许可证禁止将其用于商业目的，但突然之间任何人都可以进行试验。从这一点开始，创新来势汹汹。. The impact on the community cannot be overstated. Existing licenses prevent it from being used for commercial purposes, but suddenly anyone is able to experiment. From this point forward, innovations come hard and fast.

2023 年 3 月 12 日 – 烤面包机上的语言模型March 12, 2023 – Language models on a Toaster

一个多星期后，Artem AndreenkoA little over a week later, Artem Andreenko 让模型在 Raspberry Pi 上运行gets the model working on a Raspberry Pi。此时模型运行速度太慢而不实用，因为权重必须分页进出内存。尽管如此，这为小型化的冲击奠定了基础。. At this point the model runs too slowly to be practical because the weights must be paged in and out of memory. Nonetheless, this sets the stage for an onslaught of minification efforts.

2023 年 3 月 13 日 – 在笔记本电脑上微调March 13, 2023 – Fine Tuning on a Laptop

第二天，斯坦福发布了The next day, Stanford releases Alpaca Alpaca，它为 LLaMA 添加了指令调优。然而，比实际重量更重要的是 Eric Wang 的, which adds instruction tuning to LLaMA. More important than the actual weights, however, was Eric Wang’s alpaca-loraalpaca-lora存储库，它使用 repo, which used 低阶微调low rank fine-tuning“在单个 RTX 4090 上在数小时内”完成了此训练。 to do this training “within hours on a single RTX 4090”.

突然之间，任何人都可以微调模型来做任何事情，从而在低预算的微调项目中拉开了一场逐底竞争的序幕。论文自豪地描述了他们总共花费了几百美元。更重要的是，低等级更新可以轻松地与原始权重分开分发，从而使它们独立于 Meta 的原始许可。任何人都可以分享和应用它们。Suddenly, anyone could fine-tune the model to do anything, kicking off a race to the bottom on low-budget fine-tuning projects. Papers proudly describe their total spend of a few hundred dollars. What’s more, the low rank updates can be distributed easily and separately from the original weights, making them independent of the original license from Meta. Anyone can share and apply them.

2023 年 3 月 18 日 – 现在快了March 18, 2023 – Now It’s Fast

Georgi GerganovGeorgi Gerganov 使用 4 位量化uses 4 bit quantization在 MacBook CPU 上运行 LLaMA。它是第一个速度足够实用的“无 GPU”解决方案。 to run LLaMA on a MacBook CPU. It is the first “no GPU” solution that is fast enough to be practical.

2023 年 3 月 19 日 – 13B 型号与 Bard 实现“平价”March 19, 2023 – A 13B model achieves “parity” with Bard

第二天，跨大学合作发布了The next day, a cross-university collaboration releases Vicuna Vicuna，并使用 GPT-4 驱动的 eval 来提供模型输出的定性比较。虽然评估方法值得怀疑，但该模型实际上比早期变体更好。, and uses GPT-4-powered eval to provide qualitative comparisons of model outputs. While the evaluation method is suspect, the model is materially better than earlier variants. 培训费用：300 美元。Training Cost: $300.

值得注意的是，他们能够使用来自 ChatGPT 的数据，同时绕过对其 API 的限制——他们只是对发布在Notably, they were able to use data from ChatGPT while circumventing restrictions on its API – They simply sampled examples of “impressive” ChatGPT dialogue posted on sites like ShareGPT等网站上的“令人印象深刻的”ChatGPT 对话示例进行了采样ShareGPT。.

2023 年 3 月 25 日 – 选择您自己的模型March 25, 2023 – Choose Your Own Model

Nomic 创建了Nomic creates GPT4All GPT4All，它既是一个, which is both a 模型model ，更重要的是一个and, more importantly, an 生态系统ecosystem。我们第一次看到模型（包括 Vicuna）聚集在一个地方。. For the first time, we see models (including Vicuna) being gathered together in one place. 培训费用：100 美元。Training Cost: $100.

2023 年 3 月 28 日——开源 GPT-3March 28, 2023 – Open Source GPT-3

Cerebras（不要与我们自己的 Cerebra 混淆）使用 Chinchilla 隐含的最佳计算计划和Cerebras (not to be confused with our own Cerebra) trains the GPT-3 architecture using the optimal compute schedule implied by Chinchilla, and the optimal scaling implied by μ 参数化隐含的最佳缩放来训练 GPT-3 架构μ-parameterization。这大大优于现有的 GPT-3 克隆，并且代表了“在野外”首次确认使用 μ 参数化。这些模型是从头开始训练的，这意味着社区不再依赖 LLaMA。. This outperforms existing GPT-3 clones by a wide margin, and represents the first confirmed use of μ-parameterization “in the wild”. These models are trained from scratch, meaning the community is no longer dependent on LLaMA.

2023 年 3 月 28 日 – 一小时多模式培训March 28, 2023 – Multimodal Training in One Hour

Using a novel Parameter Efficient Fine Tuning (PEFT) technique, LLaMA-AdapterLLaMA-Adapter使用一种新颖的参数高效微调 (PEFT) 技术，在一小时的训练中引入了指令调优和多模态。令人印象深刻的是，他们仅使用 120 万个可学习参数就可以做到这一点。该模型在多模态 ScienceQA 上实现了新的 SOTA。 introduces instruction tuning and multimodality in one hour of training. Impressively, they do so with just 1.2M learnable parameters. The model achieves a new SOTA on multimodal ScienceQA.

2023 年 4 月 3 日——真人无法区分 13B 开放模型和 ChatGPTApril 3, 2023 – Real Humans Can’t Tell the Difference Between a 13B Open Model and ChatGPT

伯克利推出了Berkeley launches Koala Koala，这是一种完全使用免费数据训练的对话模型。, a dialogue model trained entirely using freely available data.

他们采取了关键步骤来衡量他们的模型和 ChatGPT 之间的真实人类偏好。虽然 ChatGPT 仍然略有优势，但超过 50% 的用户要么更喜欢 Koala，要么没有偏好。They take the crucial step of measuring real human preferences between their model and ChatGPT. While ChatGPT still holds a slight edge, more than 50% of the time users either prefer Koala or have no preference. 培训费用：100 美元。Training Cost: $100.

2023 年 4 月 15 日 – ChatGPT 级别的开源 RLHFApril 15, 2023 – Open Source RLHF at ChatGPT Levels

Open AssistantOpen Assistant启动了 launches 一个模型，更重要的是，启动了一个a model and, more importantly, a dataset用于通过 RLHF 对齐的数据集。他们的模型在人类偏好方面与 ChatGPT 接近（48.3% 对 51.7%）。除了 LLaMA，他们还展示了这个数据集可以应用于 Pythia-12B，让人们可以选择使用完全开放的堆栈来运行模型。此外，由于数据集是公开可用的，因此对于小型实验者而言，RLHF 从无法实现变为廉价且容易。 for Alignment via RLHF. Their model is close (48.3% vs. 51.7%) to ChatGPT in terms of human preference. In addition to LLaMA, they show that this dataset can be applied to Pythia-12B, giving people the option to use a fully open stack to run the model. Moreover, because the dataset is publicly available, it takes RLHF from unachievable to cheap and easy for small experimenters.