###### September 17, 2024


Meta, like many other leading tech companies, has developed its own generative AI model known as [Llama](https://llama.meta.com/). What sets Llama apart is its open nature, allowing developers to download and use it with considerable freedom, though there are some conditions to keep in mind. This approach contrasts with models like Anthropic’s Claude, [OpenAI’s GPT-4o](https://openai.com/index/hello-gpt-4o/) (the engine behind [ChatGPT](/news/ai-at-your-fingertips-chatgpt-app-now-available-for-mac-2024-07-02)), and Google’s Gemini, which are available solely through APIs.

To give developers more flexibility, Meta has partnered with cloud service providers such as AWS, Google Cloud, and Microsoft Azure to offer cloud-hosted versions of Llama. Additionally, Meta has built tools that simplify the process for developers to fine-tune and adapt the model to meet their specific requirements.

Below is a detailed description of Llama, including what it can do, the different versions it offers, and how you can use it effectively.


## What is Llama?

Llama is not merely one model; it encompasses a range of models, such as:

- Llama 8B
- Llama 70B
- Llama 405B

The most recent versions (Llama 3.1 8B, Llama 3.1 70B, and Llama 3.1 405B) were released in July 2024. These models are trained on a wide range of data sources, such as web pages in different languages, public code, files available online, and synthetic data produced by other AI systems.

The Llama 3.1 8B and 70B are compact models that can run a variety of devices from laptops to servers. The Llama 3.1 405B, on the other hand, is a large-scale model that typically requires data center hardware. While the smaller models may not be as powerful as the 405B, they offer faster performance and are optimized for reduced storage and latency.

All Llama models feature a context window of 128,000 tokens, allowing them to handle around 100,000 words (or 300 pages of text). This is roughly the length of books like Wuthering Heights or Harry Potter and the Prisoner of Azkaban. A long context window helps the model retain information from recent documents and data, reducing the risk of straying off-topic.


## What Can Llama Do?

Like other generative AI models, Llama can help with a variety of tasks, such as coding, answering simple math questions and summarizing documents in eight languages: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai. It can handle most text-based tasks, such as analyzing files like PDFs and spreadsheets, but it currently doesn’t have the ability to generate or process images — something that could change in the future.

Llama models can integrate with third-party apps, tools, and APIs to perform a variety of tasks. They are configured to use Brave Search to answer questions about current events, the Wolfram Alpha API for math and science queries, and a Python interpreter for code validation. According to Meta, Llama 3.1 can even use some tools it hasn’t been explicitly trained on, though the reliability of this feature is still uncertain.


## Where Can You Use Llama?

If you want to connect directly with Llama, the Meta AI chatbot experience is supported on platforms like Facebook Messenger, WhatsApp, Instagram, Oculus, and Meta.ai.

For developers, Llama is available for download and can be scaled to several popular cloud platforms. Meta has partnered with over 25 companies to host Llama, including Nvidia, Databricks, Groq, Dell, and Snowflake. Many of these partners provide additional tools that allow Llama to access proprietary data or run more efficiently.

Meta suggests using the smaller Llama models, specifically the 8B and 70B, for general purposes such as running chatbots or creating code. On the other hand, the larger Llama 405B is more appropriate for tasks such as model distillation, which involves transferring knowledge from a larger model to a smaller one, as well as generating synthetic data to train other models.

Developers operating applications with more than 700 million monthly users must acquire a special license from Meta to access Llama, and this will be granted at Meta's discretion.


## What Tools Does Meta Offer for Llama?

Meta has introduced several tools to improve Llama's security:

1. **Llama Guard:** A moderation framework that identifies problematic content such as hate speech, self-harm, and copyright infringement.

2. **Prompt Guard:** A tool specifically designed to protect Llama from harmful messages that seek to bypass its security measures.

3. **CyberSecEval:** A cybersecurity risk assessment suite that evaluates model security, focusing on threats such as automated social engineering and offensive cyber activities.

For example, Llama Guard can detect harmful or illegal content that is entered or generated into Llama, allowing developers to customize which categories are blocked. Prompt Guard focuses on defending against "message injection attacks" that attempt to manipulate the model. CyberSecEval offers benchmarks for assessing security risks associated with Llama models.


## Llama’s Limitations

Llama, like other generative AI models, has its limitations and potential risks. One of the main concerns is whether Meta used copyrighted materials to train Llama. If that is the case, users could face liability for any copyrighted content the model generates.

Recent reports indicate that Meta has been using copyrighted e-books for AI training, despite having received legal warnings. The company also incorporates content from Instagram and Facebook into its model training, making it difficult for users to opt-out. Meta is facing multiple lawsuits, including one from authors such as Sarah Silverman, who claim that the company has used copyrighted material without permission.

An important consideration is programming, as Llama could produce buggy or unsafe code. It is critical that developers have a human expert review any AI-generated code before implementing it in their applications.

While Meta's Llama model offers considerable flexibility and opportunities for developers, it is important to recognize the potential risks and limitations that come with it.

–

_Brought to you by Code Labs Academy – Your Leading_ [_Online Coding Bootcamp_](https://codelabsacademy.com/) _for Future Tech Innovators._


###### Image Credits: Meta

thumbnail_Llama-Use-Case_Header_16x9-1.webp

Llama-Use-Case_Header_16x9-1.webp

What is Meta Llama? A Comprehensive Look at the Open AI Model for Developers

###### September 17, 2024

While still a computer science student at the University of Waterloo, Jacob Jackson co-founded [Tabnine](https://www.tabnine.com/get-free-90-day-trial/?utm_term=tabnine\&utm_campaign=%5BUS%5D+%7C+%5Ben%5D+%7C+%5Bdesktop%5D+%7C+%5BSearch:Brand%5D+%7C+%5BTabnine%5D\&utm_source=adwords\&utm_medium=ppc\&hsa_acc=4311736126\&hsa_cam=21073527735\&hsa_grp=159292934283\&hsa_ad=713438295758\&hsa_src=g\&hsa_tgt=kwd-915660947489\&hsa_kw=tabnine\&hsa_mt=e\&hsa_net=adwords\&hsa_ver=3\&gad_source=1\&gclid=CjwKCAjw0aS3BhA3EiwAKaD2ZSi_B04PfNmu2pEw62uZL71RUlGpSu2J1k_0CDSFuIosfG5tgV4fVRoCod8QAvD_BwE), an AI coding assistant company, launching his career in AI early. In 2019, as Jackson was finishing his final exams, Tabnine raised about $60 million in venture capital and was sold to Codata. After that, he started working as an intern at [OpenAI](https://openai.com/) until 2022, when he made the decision to start a new company that would improve developer workflows.

Jackson developed [Supermaven](https://supermaven.com/), an AI coding platform that expands on his experience with Tabnine but features major technology enhancements, thanks to his inspiration from AI tool developments such as ChatGPT and GitHub Copilot. Babble, Supermaven's in-house AI model, has a million-token pop-up that allows it to analyze a lot of code at once. This makes the model more reliable in producing accurate code by allowing it to handle large inputs without “forgetting” recent data or straying from the topic.

Supermaven matches competitors like Google's Code Assist with 1 million [tokens](https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them), despite the fact that some AI coding platforms, like Magic, offer considerably larger pop-ups (Magic supports 100 million tokens). According to Jackson, Supermaven's unique selling point is its reduced latency, made possible by an innovative internal neural architecture. Additionally, the platform takes just 10-20 seconds to process a developer's code repository, learning coding rules and APIs so it can keep running quickly, even with large code bases.

The AI ​​coding tools market is expected to grow rapidly and reach $27.17 billion by 2032, [according to forecasts](https://www.polarismarketresearch.com/press-releases/ai-code-tools-market). More than 1.8 million GitHub Copilot users are among the many developers who have already integrated AI into their processes. But the sector has many obstacles to overcome, particularly when it comes to intellectual property and data privacy. Fearful of disclosing PIN, many companies are hesitant to use AI coding tools. Additionally, copyrighted code has been known to be regurgitated by AI models in some cases, putting developers in legal jeopardy if they include it in their work without realizing it. .

Jackson emphasized [in a talk with Techcrunch](https://techcrunch.com/2024/09/16/ai-coding-assistant-supermaven-raises-cash-from-openai-and-perplexity-founders/) that while Supermaven retains data for a brief period to maximize system performance, it does not use consumer data to build its models. Reducing exposure to dangerous content during training, he emphasized that the focus was on publicly accessible code, but he did not completely rule out the possibility that Babble would be taught on IP-protected code.

Supermaven has grown significantly in popularity despite these obstacles, with over 35,000 developers using the platform. Since the company launched in February, its user base has tripled and many of those customers are signing up for premium plans. The company's annual recurring revenue reached $1 million.

Due to the success of the platform, investors are taking note. Recently, Supermaven revealed that the company had completed its first external investment round, raising $12 million from angel investors and Bessemer Venture Partners, which included Perplexity and OpenAI co-founders John Schulman and Denis Yarats. The money will be used to improve Supermaven's text editor, which is currently in beta, and to hire more developers, setting the company up for long-term success in the fast-growing field of computer coding AI.


thumbnail_charlesdeluvio-pjAH2Ax4uWk-unsplash (1) (1).webp

charlesdeluvio-pjAH2Ax4uWk-unsplash (1) (1).webp

Supermaven Gains Investor Support: AI Tool Optimizing Developer Workflows


###### September 13, 2024


Meta is [updating](https://about.fb.com/news/2024/04/metas-approach-to-labeling-ai-generated-content-and-manipulated-media/) its approach to labeling content that has been edited or modified by AI tools on Instagram, Facebook, and Threads. The “AI Insights” label, which previously appeared directly below a user’s name, will now move to the posts menu for content that has been modified using AI.

The company said the label will continue to appear on content that Meta identifies as being entirely AI-generated. While the label may be less visible on content that has been edited using AI tools, it will still be clearly visible on content that has been created entirely with AI input.

Meta has announced that it will offer information on whether content is labeled based on industry-standard signals or user disclosure. This change, which will be implemented next week, aims to “better reflect the scope of AI used in content” across all of its platforms.

By reducing the prominence of the AI ​​information label, there is concern that users could be more easily fooled by AI-edited content, particularly as editing tools become increasingly sophisticated.

This isn’t the first time Meta has changed the way it labels AI-related content. In July, the company changed the label from “Made with AI” to “Information from AI” after receiving complaints from photographers who felt the label was incorrectly applied to real photos. Meta admitted that the original wording was confusing and could have been interpreted to mean that certain images were entirely AI-generated, even if they had only been edited with AI tools.

—

_Brought to you by Code Labs Academy – Your Leading_ [_Online Coding Bootcamp_](https://codelabsacademy.com/) _for Future Tech Innovators._


###### Image Credits: Meta

thumbnail_meta-ai-web.webp

meta-ai-web.webp

Meta Changes AI Labeling on Social Platforms for Edited and AI-Generated Content

###### September 13, 2024

Though opinions on whether Rajeev Behera's new HR firm, [Every](https://www.every.io/), is a high-risk venture or a brilliant idea are varied, the startup is making waves. For small firms, the all-in-one platform integrates payroll, onboarding, and expenditure management. This market is already congested with established players like Gusto, Rippling, and Deel as well as up-and-coming companies like Mercury and Brex. However, Behera's backers appear to be certain that his strategy is unique. Recently, Every raised $22.5 million in a Series A fundraising round headed by Okta Ventures, Base10 Partners, Y Combinator, and Redpoint Ventures.

Behera's emphasis on early-stage firms and distinctive offering strategy are what distinguish his vision. Every provides free incorporation services to small firms, and then they set up business bank accounts and provide crucial back-office support. Interchange fees and SaaS fees for extra modules like accounting generate revenue.

[Behera explained](https://techcrunch.com/2024/09/12/why-y-combinator-companies-are-flocking-to-new-banking-hr-startup-every/) that Every has developed advanced tools for payroll, banking, and expense management. To attract founders early on, they are now offering free incorporation services. Once onboarded through a personalized session, startups gain access to a comprehensive system covering banking, payroll, HR, bookkeeping, taxes, and state compliance. Additionally, founders benefit from a dedicated Slack channel where they can connect and share insights with other entrepreneurs.

After completing their incorporation using Every, customers usually stick with it instead of using competitors' platforms like Rippling. With tools meant to promote their growth during the first five years, Every's concentration is on companies with less than 200 people.

Behera's approach is unique in that he anticipates his most profitable clients will eventually require more services from Every—at least until Every develops into a growth-stage fintech that can assist bigger companies.

Behera's relationship with [Y Combinator](https://www.ycombinator.com/) (YC) has been a vital component of Everybody's development. Behera, an experienced entrepreneur, made use of the accelerator's network, and everyone was a part of YC's Summer 2023 cohort. Reflektive, his prior business endeavor, was purchased in 2021 following a capital raise of more than $100 million. In addition, YC advisor Surbhi Sarna, his wife, is also a successful entrepreneur. Because of this, about half of Every's clientele came via YC's extensive startup network.

Even though Rippling, Gusto, and Deel—three other YC-backed startups—compete with them, Every's close ties to the YC ecosystem are a significant advantage. This has attracted investors, along with Behera's hands-on approach to creating the platform's components from the ground up.

Behera hired more staff after personally overseeing customer sales, support, and product design for the first 50 clients of the business, with the help of Redpoint's Alex Bard, who oversaw the Series A fundraising round. After reading glowing testimonials from firms that were already use Every, other venture capitalists did the same.

In less than two weeks, the Series A money was secured without Behera actively seeking new investors. Although the company's worth was not made public, estimates place it at $112.5 million. The additional funding will be used to support the free incorporation and onboarding services that are driving Every's early success, as well as to grow the engineering staff, which now numbers roughly 20 people.


What is Meta Llama? A Comprehensive Look at the Open AI Model for Developers

Supermaven Gains Investor Support: AI Tool Optimizing Developer Workflows

Meta Changes AI Labeling on Social Platforms for Edited and AI-Generated Content

HR Startup Every Secures $22.5M Series A Funding and Expandes Engineering Team in Just Two Weeks