🤖 AI Brief #17: Is Generative AI a dud? The big debate happening right now.

Today is Monday, August 21, 2023.

Welcome to issue 17 of our AI Brief! We kick off this issue with a provocative yet important question to ask: how overhyped is generative AI?

A thoughtful piece by Gary Marcus is currently at the center of attention right now in the AI community. After months of seeing what feels like breathless advancements in AI, he wants us to consider the opposite: is it all just going to lead to disappointment?

An important question I ask is how we can stay on top of what’s truly meaningful in the world of AI while filtering out the hype that leaves us fatigued. This is a key motivation for why I avoid adding long laundry lists of news developments to this newsletter.

Less is more. If there’s important news to feature — it’ll end up here. Otherwise I try to stay away from the unneeded hype.

And of course, this is just one humble person’s opinion.

In this issue:

• 🧠 The big debate: what if generative AI turns out to be a dud?

• 🖼️ AI-created art isn’t copyrightable, federal judge rules

• 👊 Meta prepares to launch its own open-source code generation AI model

• 📦️ Amazon moves to summarize product reviews with AI, and other quick scoops

• 🧪 The latest science experiments, including the first LLM agent benchmark

Did someone forward this to you? Subscribe here!

🧠 The big debate: what if generative AI turns out to be a dud?

Driving the conversation within AI circles is a thought-providing piece from entrepreneur and neuroscience professor Gary Marcus.

He asks a simple question: what if generative AI is set up to disappoint us?

Outside of writing text and code, Marcus argues, “other potential paying customers may lose heart quickly” as the limits on generative AI become clearer to the professional world.

• The love affair with ChatGPT is wearing off, Marcus points out, as a growing chorus of voices become disillusioned with its limitations.

• OpenAI’s $29B valuation may be hard to justify, and other startups raising at lofty valuations may find themselves in hot water.

• And as numerous startups broadly claim they’re AI companies, the AI hype bubble risks distorting the real work that remains to be done in the field.

AGI is unlikely to be imminent, no matter what the optimists would have you believe, Marcus writes in a separate blog post. The long march towards improving artificial intelligence systems is still rife with unsolved challenges, and LLMs are unlikely to be a pathway to AGI.

A personal thought: as a newsletter writer, I’ve always shied away from the hyped-up proclamations of “ChatGPT is yesterday’s news” and “AI Agents just made Microsoft obsolete,” etc. Deep down, I believe AI can be big, but a quick survey of friends show that we’re personally experiencing some AI fatigue as well.

Venture capitalist Benedict Evans sums up the feeling of the moment quite well:

🖼️ AI-created art isn’t copyrightable, federal judge rules

A federal judge has upheld a finding from the U.S. Copyright Office that AI-created art isn’t copyrightable. Existing copyright claw cannot “protect works generated by new forms of technology operating absent any guiding human hand,” U.S. District Judge Beryl Howell wrote in her decision.

At the core of the finding is a foundational rule: “Human authorship is a bedrock requirement” for to grant a copyright.

• This tracks with previous legal rulings, including one that determined a photograph captured by a monkey couldn’t be granted copyright. Copyright law is meant for “human individuals to engage in” creation, the judge in that case ruled.

However, one exception remains: works created with the help of AI could still be eligible for copyright, if a human “selected or arranged” it in a “sufficiently creative way that the resulting work constitutes an original work of authorship,” the copyright office said.

👊 Meta prepares to launch its own open-source code generation AI model

Meta’s open-source wins will likely mount as it readies another model for the public – this time focused on code generation. Meta’s model, dubbed Code Llama, could pose a threat to paid coding assistants like GitHub’s Copilot, which currently uses OpenAI’s AI models to power its suggestions.

Why this matters:

• Coding assistants have been one of the areas of AI tools to see rapid adoption, as increasingly more powerful LLMs and an eager user base of software developers have witnessed significant leverage in the last year.

• In particular, an open-source model could be highly attractive to companies focused on security: an enterprise could theoretically take Code Llama, fine-tune it on their own code base, and keep the model completely walled off.

One likely casualty of Meta’s Code Llama release could be all of GitHub’s startup competitors, who don’t have GitHub’s reach, Microsoft backing, and head start in user base. And they now have to contend with a free open-source model likely to see rapid adoption.

🔎 Quick Scoops

Advertisers are exploring generative AI tools to cut costs and increase productivity, but overall remain cautious about security and copyright risks. (Reuters)

Academic journals are fighting an influx of shoddy research papers largely written by AI. Concerns about harming their own credibility abound as major journals rush to put in place better safeguards and policies around the use of AI tools. (Decoder)

Google is building an AI to offer life advice, and their DeepMind unit is testing focused AI models designed to help users navigate intimate questions. (CNBC)

Amidst the Hollywood strikes, 96% of entertainment companies plan to boost generative AI spend. This pace of planned AI investment is worrying to many parties. (Forbes)

Amazon rolls out product review summaries made by generative AI. Starting first on mobile, this is designed to help users cut through the noise of browsing product user-written reviews. (AP News)

🧪 Science Experiments

AgentBench: the first benchmark designed to evaluate LLM-as-Agent

• Benchmarks to evaluate LLM performance on focused areas exist, but none to date look at how LLMs can perform as agents.

• AgentBench offers 8 distinct environments to provide a more comprehensive evaluation of the LLMs' ability to operate as autonomous agents in various scenarios.

• See the project page here.

Dual-Stream Diffusion Net for Text-to-Video Generation

• Videos created from a text-to-video generative model carry flickers and artifacts. By separating out the process into two separate diffusion streams - one for video, and one for motion – the end result is significantly improved.

• See the paper here. See their example videos here.

TeCH: Text-guided Reconstruction of Lifelike Clothed Humans

• Despite recent research advancements in reconstructing clothed humans from a single image, accurately restoring the "unseen regions" with high-level details remains an unsolved challenge that lacks attention.

• TeCH produces high-fidelity 3D clothed humans with consistent & delicate texture, and detailed full-body geometry.

• See their paper here.

👋 How I can help

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, EvenUp, Tome, Twelve Labs and more to help make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at [email protected] 

As always — have a great week!