OpenAI’s “Orion” shoots for the stars, xAI’s Colossus flexes its muscles, and Ligo unfolds a new era in bioinformatics — the top 3 AI news stories of the week
Our latest AI Digest covers the most intriguing developments in the world of artificial intelligence for the week. Valeriy Selitskiy, AI enthusiast and Software Engineer, comments on key stories.
#1 — OpenAI’s “Orion” shoots for the stars, leaving Q* to wither on the vine
In a celestial leap forward, OpenAI is reportedly cultivating two new AI powerhouses: “Strawberry” (also known as Q*) and the more ambitious “Orion.” Strawberry’s math and problem-solving skills are formidable, but it’s Orion that is poised to eclipse GPT-4’s capabilities. OpenAI is adopting a unique approach: using Strawberry to supply Orion with high-quality synthetic data, potentially reducing those pesky AI hallucinations. As Orion continues its stellar development before a potential end-of-year launch date, Strawberry may find its way into ChatGPT as soon as the fall of 2024, proving that even in the AI world, some fruits ripen faster than others. It’s worth noting that these details come from unnamed sources and should be taken with a grain of salt. As with many rumors in the fast-paced world of AI, the full picture may emerge only when OpenAI decides to make an official announcement.
#2 — xAI’s Colossus: a titan of computing power with an appetite for megawatts
Elon Musk’s xAI has unleashed “Colossus,” a behemoth 100k H100 training cluster that’s living up to its mythological namesake. Created and brought online in just 122 days, this digital titan is flexing its muscles as the world’s most powerful AI training system. But with great power comes great electricity bills — Colossus is expected to consume up to a whopping 195 megawatts of power at peak usage times (and use up to a million gallons of water per day to cool its servers). Since xAI plans to double its capacity in the coming months, one can’t help but wonder if they’ll need their own power plant. This technological marvel promises to supercharge advancements in autonomous vehicles, robotics, and xAI’s Grok language model, proving that in the world of AI, bigger really might be better.
#3 — Ligo unfolds the future of bioinformatics with open-source AlphaFold3
In a move that is sure to send ripples through the scientific community, Ligo has released an open-source implementation of AlphaFold3, unfolding a new chapter in biomolecular structure prediction. This PyTorch-based marvel, available under the Apache 2.0 license, isn’t just a carbon copy — Ligo’s team is ironing out discrepancies between the original pseudocode and deep learning literature, enhancing both efficiency and performance. Partnering with AdaptyvBio for expanded ProteinFlow capabilities, and Basecamp Research for high-quality sequences, Ligo is stitching together a tapestry of innovation to dramatically accelerate the mapping of protein structures. Instead of requiring the entire duration of PhD studies to map a single structure, AlphaFold3 can provide a prediction in minutes. As beta testing continues, the scientific community eagerly awaits the full unveiling of this game-changing tool, which promises to reshape the landscape of bioinformatics.
.png?auto=webp)
