We Are Accelerating

Last week was arguably one of the craziest, most exhilarating, and most jam-packed weeks of technological progress we’ve seen in quite a while. And it wasn’t simply the unprecedented quantity of new products, innovations, and events; it was also the scale and breadth of technological progress we witnessed all in just a few days. If you blinked, you may have missed something. So we here at The Technology Brother thought it would be prudent to give you a recap - a walkthrough of all the technological advancements we saw last week. Buckle up.

Extropic

Early in the morning on Monday, the 11th, Guillaume Verdon (also known as Beff Jezos on X) and his team at Extropic finally unveiled the groundbreaking technology they were working on. In a “Litepaper,” as well as a sit-down interview with Garry Tan, Verdon and CTO Trevor McCourt revealed their new Extropic chip. The chip is designed to harness thermodynamics, via the fluctuations in energy at the level of electrons, enabling an AI accelerator that is many orders of magnitude faster and more energy efficient than digital processors. You can read all about it in their Litepaper, here.

Cognition’s Devin SWE

The very next day would see another incredibly impactful release: that of Cognition Lab’s “Devin” - “the first AI software engineer.” For the whole day, Devin was the talk of X. Devin represents a breakthrough as potentially the world's first fully autonomous AI software engineer.  It sets a new benchmark in AI's capability to perform complex software engineering tasks without human intervention, demonstrating proficiency across a spectrum of developer tasks—from debugging and deploying apps to training AI models. Devin also significantly outperformed previous AI models in the SWE-bench coding benchmark by solving over 13% of real-world GitHub issues autonomously, a leap from the prior best of under 5%.

Truffle

That same afternoon, we saw the release of Truffle-1, “an AI inference engine designed to run open source models at home, on 60 Watts.” It’ll be able to run Mixtral models at 22+ tokens/s without speculative decoding, which is 2x faster than GGML on the Nvidia Orin architecture. Beyond its technical prowess, Truffle-1’s appeal lies in its aesthetic beauty, reminiscent of the “Wetware” from Ex Machina, as well as its versatility, connecting via BLE, WiFi, or USB-C to enhance devices with AI capabilities, suggesting a future where applications evolve from traditional software to intelligent, AI-augmented solutions.

Figure Demo

And to think, the week was just getting started. Wednesday was a day for the history books. Let’s start with Figure’s groundbreaking new demo video, which it released on X. Figure plans to “deploy autonomous humanoid robots on a global scale to solve challenges within the labor economy.” And their demo on Wednesday showed off their new partnership with OpenAI, demonstrating a Figure robot adapted with access to GPT-4, giving it stunning new high-level visual and language intelligence. Its ability to recognize its environment, reason about it, and interact with it is simply stunning. The demo is a must-watch if you haven’t seen it already.

Claude Haiku

Then, there was the release of Anthropic’s new Claude Haiku model. Anthropic had already unveiled the Claude 3 foundation model family, with AWS subsequently announcing the general availability of Claude 3 Sonnet on Amazon Bedrock. This new announcement introduced Claude 3 Haiku's availability on Amazon Bedrock, and showcased its position as the fastest and most compact model within the Claude 3 lineup, optimized for near-instantaneous response times. Such capabilities will be ideal for applications requiring rapid and precise generative AI abilities.

DeepMind’s SIMA

Of course, Google and DeepMind also had to get in on the fun. On the 13th, they released a brand-new research paper at the cutting-edge of AI agents. SIMA, short for Scalable Instructable Multiworld Agent, is a generalist AI agent for 3D virtual settings that can “follow natural-language instructions to carry out tasks in a variety of video game settings.” SIMA marks a significant leap in AI capabilities, presenting an agent that can perceive, understand, and interact with a variety of virtual environments without needing game source codes or bespoke APIs. Instead, it operates using visual inputs and natural language instructions to control game characters, showcasing its adaptability across 600 basic skills like navigation and object interaction. SIMA brings to mind a future where AI can execute complex, strategic tasks autonomously. Incredible stuff.

Cerebras

Meanwhile, Extropic would not be the only chip released this past week. On Wednesday, Cerebras Systems, one of Nvidia's most prominent competitors, unveiled the "Wafer Scale Engine 3," the third generation of its AI chip and the world's largest semiconductor. It is truly massive, just look at the thing. The WSE-3 has the computational capacity of 125 petaFLOPS (one petaFLOP equals one quadrillion operations per second). Nearly the same size as a 12-inch wafer, it contains 4 trillion transistors at 5 nanometers in size. TSMC is manufacturing the WSE-3.

Reflect Orbital

To cap off a crazy Wednesday, there was one more thing that almost went unnoticed, but definitely worth mentioning here - and that was the announcement of a new, fascinating venture: Reflect Orbital. The company aims to “develop a constellation of revolutionary satellites to sell sunlight to thousands of solar farms after dark.” It’s a bold idea. What if you could reflect sunlight from up in space down to the ground where night had already fallen? On August 31st, they completed their final on-earth testing and are currently designing their first satellite.

Starship!

On Thursday came easily the biggest, most incredible event of the week - SpaceX finally launched its 3rd integrated flight test of Starship, and wow, did it deliver. The launch was a massive success and an astonishing spectacle. Starship successfully lifted off from the pad in Boca Chica, flawlessly executed a hot-stage separation, and put Starship into an orbital velocity. There were small hiccups; notably, the Super Heavy Booster came down hard and exploded just above the Gulf of Mexico, meanwhile Starship experienced some control issues once in space and re-entered the atmosphere in a roll, causing it to be lost. Altogether, though, the launch was a total success and brings humanity one step closer to the Moon, and Mars.

Anduril ALTIUS

Lost amongst the hype and attention of the Starship launch was an interesting announcement out of Anduril. On the 14th, Anduril announced the ALTIUS-700M, a groundbreaking advancement in loitering munitions with unparalleled payload capacity, superior loitering time, and sophisticated autonomy for executing coordinated strikes. The technology, tested successfully in Utah, is a huge leap for Anduril, allowing for precise, high-impact attacks against armored targets and infrastructure with a range of up to 100 miles and 75 minutes of flight time. Enemies of the West, beware.

Grok-1 Open-Sourced

The rest of the weekend was comparatively slower, as everyone waited with bated breath for whether or not Elon Musk would come through on his promise. “This week, @xAI will open source Grok,” he tweeted on Monday. Finally, on Sunday, he delivered. In a blog post linking to GitHub, xAI announced they had open-sourced the raw base model checkpoint from the Grok-1 pre-training phase, trained from scratch using a custom training stack on top of JAX and Rust in October 2023. The model was trained on a large amount of text data, and not fine-tuned for any particular task, such as dialogue. It’s a 314B parameter Mixture-of-Experts model with 25% of the weights active on a given token. And in a great win for open source, the weights and the architecture were released under the Apache 2.0 license. True to his word, Elon delivered.

From start to finish, last week was an absolute rollercoaster of technological progress. And already, not to be outdone, this week is looking like it’ll be more of the same. Is this what acceleration looks like?