AI Week In Review 24.04.20
Boston Dynamics' new Atlas, Meta Llama 3 8B & 70B, Reka Core, Code Qwen1.5-7B, Aiola fast TTS, Adobe Gen AI in Premier Pro, Open AI Batch, Limitless Pendant, VASA-1, Stanford's '24 AI report.
AI Tech and Product Releases
There was a slew of new AI model releases, including Meta’s Llama 3 release, which we covered in depth in “AI models unleashed! Llama 3 launches!”
Meta’s Llama3 8B and 70B models have raised the bar for everyone, in many ways. Llama 3 70B’s near-GPT-4-level performance obsoletes many prior AI models, and Llama 3 8B is better than GPT-3.5 and SOTA for its size. More is coming from Meta - a larger LLM and multi-modality.
Reka announced Core, a very solid multi-modal LLM: “Core is comparable to GPT-4V on MMMU, outperforms Claude-3 Opus on our multimodal human evaluation, and surpasses Gemini Ultra on video tasks.”
Alibaba’s Qwen team announced Code Qwen-1.5, a 7B coding model with 64K context. Code Qwen-1.5-7B-chat model gets 83.5 on HumanEval and 23.2 on LiveCodeBench, which is best-in-class for its size and SOTA for an open AI model you can download.
Mistral announced more details on its 8x22B AI model, including 64K context, native function-calling, and good multi-lingual capabilities.
Israeli startup Aiola has launched a text-to-speech model, that improved on OpenAI’s Whisper by being both better and 160x faster.
Adobe announced Premiere Pro is integrating Generative AI for video. Adobe announced a new set of generative AI video editing tools for Premiere Pro, including the ability to extend clips, add or remove objects, and generate B-roll footage. They are also planning on integrating third-party AI models from OpenAI, Runway, and Pika Labs. No release date for these upcoming features beyond ‘this year.’
OpenAI introduced a Batch API, a new asynchronous service that allows users to submit jobs that can take hours to complete. Users of the API get a 50% discount on regular completions and much higher rate limits (250M input tokens enqueued for GPT-4T), versus API queries demanding real-time responses.
Boston Dynamics presented their all-new all-electric Atlas robot in a short but compelling intro video. Rising up like a robot in a Terminator movie struck some as ‘creepy’ and led one journalist to opine “Maybe I don’t want a Rosey the Robot after all.”
Limitless' $99 AI wearable to promises to remember your meetings and everything else. The company behind the Rewind pendant have re-named their company to Limitless, and are selling a $99 pendant. (Actually pre-selling, it is shipping Q4 2024.) The pitch:
Pendant is an elegant, lightweight wearable that remembers what you say throughout the day, from in-person meetings, impromptu conversations, and personal insights.
This sounds like the practical yet humble AI device that doesn’t try to replace your smartphone that the Humane Pin should have been. Props to Limitless for pitching it right after Humane Pin launch.
Nothing has announced that it plans to more deeply integrate ChatGPT with its smartphones and earbuds. This sounds great, but the cynical critic in me will point out I can get voice ChatGPT on my Android by using the ChatGPT app directly.
Top Tools & Hacks
Meta’s Meta.ai assistant, now based on Llama 3 is our top tool of the week. It’s got Llama 3 70B (on par with Claude Sonnet) under the hood, and some practical features that I’ve already found make it a useful AI chatbot:
It has search integration, so it can answer news and current event queries. Zuck in interviews mentioned Google link, but my meta.ai used Bing. Either way, it works.
It can do Real-time image generation. It can generate images faster than you can type. On the downside, it’s not as high-quality as some alternatives.
Embedded in Meta social platforms: I just wasted several minutes doing 1980s music trivia with the AI bot; got 7 for 7. Ask “Quiz me on X” and it will. Social media users will likely use this AI as a fun time-waster.
I stated in my prior Llama 3 article that Meta’s social media platforms shouldn’t be underestimated as a powerful distribution channel for AI. Meta knows this and is taking advantage of it, providing AI to masses of people through Meta.AI and their other AI tools.
AI Research News
Another busy week on all fronts of AI. Here are our AI research highlights for this past week, shared in “AI Research Roundup 24.04.19.”
VASA-1: Lifelike Audio-Driven Talking Faces Generated in Real Time. VASA-1 has impressed many with their incredibly life-like talking head videos, getting headlines like “Microsoft's AI app VASA-1 makes photographs talk and sing with believable facial expressions.”
StableAudio2.0: Generating Long-form Music with Latent Diffusion
Megalodon: Unlimited Context Length LLMs
ResearchAgent: Automating Research Idea Generation with LLM Agents
GoEX: Towards a Runtime for Autonomous LLM Applications
Improving LLMs via Imagination, Searching, and Criticizing
A trend to take note of is diffusion transformers; they are being used in multiple generative audio-video AI models. They are all different specific architectures, but Sora, StableAudio2.0, VASA-1, and other similar AI models have all used the diffusion transformer combination in their AI models.
AI Business and Policy
OpenAI Expands with Japan Office, announcing their first office in Asia and also that “we’re releasing a GPT-4 custom model optimized for the Japanese language.” OpenAI is both tapping into Japanese talent, but also wanting a local presence in an important market.
On a related note, Japan wants to be a bigger player in AI. It’s not a surprise then that Japan will fund KDDI, four others to build AI supercomputer, a $470 million spend, as part of an effort to support home-grown AI infrastructure development.
Japan may have a specific agenda for AI. The BBC asks, “Can AI help solve Japan’s labour shortages?” Japan is aging and in demographic decline, with their labor force expected to decline by 12% from 2022 to 2040. The article mentions AI-powered cooking robots, AI for crop inspection, and AI tutor assistants. The Japanese have a more positive view on AI than any other nation, perhaps because they welcome AI doing human work.
Microsoft's OpenAI partnership could face EU antitrust probe. Sources say the “European Union antitrust regulator had decided not to investigate the partnership under EU merger rules, but that Microsoft could still face an antitrust investigation.”
DARPA just held the world’s first dogfight between AI- and human-piloted F-16s! The DoD’s Office of Advanced Research Projects aka DARPA shared a video on their Air Combat Evolution (ACE) program progress, and showed a training air battle between an AI-piloted X-62A robotic fighter jet and a human-piloted F-16 fighter jet.
AI is a “fundamental change in the news ecosystem” says media expert David Caswell. What he is seeing in journalism and media is what we can expect in many professional fields - the professional becomes the manager of an AI workflow tool:
These are tools, you bring your news gathering on the left side: your PDF, transcripts, audios, videos.. roughly. It helps you do things like analysis, summaries, turn into scripts, audios. They're orchestrated by the tool.
What the journalist is doing is coordinating the tool, verifying the content all the way through to the end, and editing. The job becomes using the tool, like an editorial manager of this AI tool.
“… more media will probably be created and originated and sourced by machines. So machines will do more gathering in a lot of journalism, will do more of the producing, the audio, the video and the text, and will create the kind of experiences of consumption that consumers have.” - David Caswell
AI Opinions and Articles
Stanford University published its annual AI Index report for 2024. This is a hefty, data-filled 500-page report that covers AI technology, applications, policy and societal impact topics in detail.
It defies easy summarization, but Venture Beat headlined their report on it with AI surpasses humans on several fronts, but costs are soaring. Stanford HAI reports’ top ten take-aways are:
AI beats humans on some tasks, but not on all.
Industry continues to dominate frontier AI research.
Frontier models get way more expensive. “For example, OpenAI’s GPT-4 used an estimated $78 million worth of compute to train, while Google’s Gemini Ultra cost $191 million for compute.”
The United States leads China, the EU, and the U.K. as the leading source of top AI models.
Robust and standardized evaluations for LLM responsibility are seriously lacking.
Generative AI investment skyrockets. “ funding for generative AI surged, nearly octupling from 2022 to reach $25.2 billion.”
The data is in: AI makes workers more productive and leads to higher quality work.
Scientific progress accelerates even further, thanks to AI. They mention AlphaDev, GNoME, and other AI applications for science.
The number of AI regulations in the United States sharply increases.
In 2023, there were 25 AI-related regulations, up from just one in 2016. Last year alone, the total number of AI-related regulations grew by 56.3%.
People across the globe are more cognizant of AI’s potential impact—and more nervous.
A Look Back …
In March 2023, I discussed how AI was getting constantly cheaper in my article “The Price of AI Plummets.” Back then, I noted that:
The latest OpenAPI pricing is 0.002/1,000 tokens for the latest and greatest GPT-3.5 versus 0.06/1,000 tokens for GPT-3 beta in 2021. It’s a 30X price reduction in 2 years for a model that is much better.
By November, pricing for gpt-3.5-turbo-1106 was: $1/1M input and $2/1M output.
Now, OpenAI pricing for gpt-3.5-turbo-0125 is: $0.50 /1M input $1.50 /1M output.
With the release of Llama 3, pricing for inference takes another dive. Anyscale pricing for access to Llama 3 APIs:
Llama 3 8B (on par with GPT-3.5): $0.15 / Million tokens
Llama 3 70B (significantly better than GPT-3.5): $1.00 / Million tokens
So, AI builders can either take a 3-10x price reduction for the same performance, or move up to a higher-quality AI model for the same price. Llama 3 has raised the bar on performance and lowered the boom on price.
Last year, I predicted a 100x reduction in the price of AI in 5 years. That is likely a conservative prediction, as AI is getting cheaper at a much faster pace.