AI Roundup April 2022: DALL-E 2 Lands, Google's PaLM Hits 540B, and Adept Bets on Agents

TL;DR

April 2022 was the month generative AI stopped being a research curiosity and started looking inevitable. OpenAI dropped DALL-E 2 and showed the world what a serious text-to-image model could do. Google countered with PaLM, a 540 billion parameter monster that turned scaling into a spectacle. And a new startup called Adept raised a Series A on a bet that the real prize is not a chatbot but an AI that actually operates your software for you.

OpenAI Unveils DALL-E 2 and Resets the Bar for Text-to-Image

On April 6, OpenAI announced DALL-E 2, the successor to the original DALL-E, and it was the kind of demo that makes people who do not follow AI suddenly start following AI. Type a sentence, get a high-resolution image that actually looks like the sentence. It could combine concepts, attributes, and styles, do inpainting (edit part of an existing image from a prompt), and produce variations on a theme that held together.

Technically, the jump came from pairing a CLIP-style understanding of images and text with a diffusion-based generator, a recipe that would go on to define the next two years of this space. The model was not open, and access started as a tightly gated waitlist with usage controls and content filters, which set off the first round of the now-familiar fight over who gets to use these tools and what they are allowed to make.

Why it matters to builders: this is the moment the text-to-image gold rush started. If you have been running Stable Diffusion on your own box, fine-tuning LoRAs, or wiring image generation into a product, the lineage traces straight back to this announcement. DALL-E 2 proved the demand was real and the quality was finally good enough to ship.

Google's PaLM Scales to 540 Billion Parameters

Days earlier, on April 4, Google introduced PaLM, the Pathways Language Model, a dense 540 billion parameter transformer trained on 780 billion tokens. The accompanying paper landed on April 7. The headline number was huge, but the more interesting story was the infrastructure: Google trained it across 6,144 TPU v4 chips using its Pathways system, hitting a hardware FLOPs utilization that was a record for a model that size.

On the results, PaLM set new state-of-the-art marks across a wide battery of language tasks and showed the kind of emergent jumps on reasoning benchmarks (think multi-step word problems explained via chain-of-thought style prompting) that fueled the whole "scale unlocks new capabilities" argument.

The catch for the rest of us: PaLM was a closed research artifact. No weights, no API, no homelab download. So while it reshaped the conversation about what was possible, it mostly served as a flex and a research north star rather than something you could build on directly.

Adept Raises a Series A to Build AI That Uses Your Software

On April 26, a young San Francisco startup called Adept raised a $65 million Series A led by Addition and Greylock. The founding team carried serious pedigree, including authors of the original Transformer paper, which got people paying attention fast.

Adept's pitch was different from the prevailing chatbot energy of the moment. Instead of a model that talks, they wanted a model that does: an AI teammate that takes a natural-language instruction and then actually drives the tools you already use, clicking through web apps and APIs to get multi-step work done. In April 2022, well before "AI agents" became the phrase everyone overused, this was an early and concrete bet on agentic action over conversation.

For builders, this is worth marking as a milestone in the agent timeline. The thesis that the unit of value is completed tasks, not generated text, is one a lot of the ecosystem would eventually converge on.

Key Takeaways

• Generative image AI went mainstream. DALL-E 2 turned text-to-image from a niche demo into a category, and everything from Stable Diffusion tooling to image features in your favorite apps follows from this moment.
• Scale was still the loudest story. PaLM's 540 billion parameters and its emergent reasoning gains kept the "bigger is better" narrative front and center, even though the model stayed locked behind Google's walls.
• Closed beat open this month. Both flagship releases were gated or fully private, a reminder that in early 2022 the frontier was not something you could self-host.
• The agent thesis got funded early. Adept's Series A put real money behind AI that operates software, staking out agentic work long before it was a buzzword.
• Pedigree moved capital. Transformer co-authors starting a company was enough to pull a sizable round, a sign of how concentrated the talent and the hype were.