DeepMind's AlphaProteo AI is outpacing years of scientific research

Designing proteins from scratch has long been a scientific puzzle. Now, Google DeepMind believes it's one step closer to solving this problem.

 By the CogX R&I team

September 11, 2024

Designing proteins from scratch is incredibly challenging. Some scientists even say it's one of the toughest problems humanity has ever faced. Now, Google DeepMind thinks it has made a significant breakthrough with AlphaProteo, its new AI system. Like a master key maker, AlphaProteo can create custom proteins that precisely bind to target molecules, potentially revolutionising drug design and disease research.

What makes it so powerful? AlphaProteo has learned the intricate dance of molecular binding from vast amounts of protein data. Given a target molecule and preferred binding locations, it generates a custom-fit protein designed to latch on at those specific sites. This could dramatically speed up research across a broad spectrum of fields, from drug development to crop pest resistance.

AlphaProteo even passes the "VEGF-A test":

• It's the first AI tool to successfully design a protein binder for VEGF-A, a protein associated with cancer and diabetes complications.

• Across seven target proteins tested, AlphaProteo achieved binding affinities up to 300 times better than existing methods.

• For the viral protein BHRF1, 88% of AlphaProteo's candidate molecules bound successfully in wet lab testing.

Not a universal solution yet. While impressive, AlphaProteo has its limits. It was unable to design successful binders against TNFɑ, a protein linked to autoimmune diseases like rheumatoid arthritis. The team acknowledges there are still many bioengineering challenges to overcome in the research and development process.

Now read the rest of the CogX Newsletter

Reflection 70B model maker breaks silence amid fraud accusations

 Matt Shumer and his team at HyperWrite stirred up the AI community last week with Reflection 70B, an open-source language model that promised to outperform even OpenAI's GPT-4. But the excitement quickly turned to scepticism as researchers struggled to replicate its impressive benchmarks.

This fine-tuned version of Meta's Llama architecture claimed to use an innovative "reflection" mechanism to check its own work, potentially solving the hallucination problem plaguing many AI models. However, third-party evaluators have raised serious questions about its performance and origins.

What went wrong: Days after the initial announcement, independent analysts reported significantly lower scores on key benchmarks like MMLU. Shumer claimed a problem during the upload process to Hugging Face might be responsible, but this explanation has left many unsatisfied.

The controversy deepens: Some critics have accused Shumer of fraud in the AI research community, while others speculate the model might be a wrapper for Anthropic's proprietary Claude 3. Shumer has apologised for getting "ahead of himself" but hasn't yet provided corrected model weights.

Research: Compute North vs. Compute South. Who controls the most powerful AI chips?

A new Oxford study reveals that powerful AI chips, essential for developing and running AI applications, are highly concentrated in a few countries, primarily the US and China.

Approximate locations of public cloud regions and the most advanced GPU type available in each region. (Source: Lehdonvirta, V. et al., 2024)

The US has a significant advantage in AI chip distribution, particularly with the powerful Nvidia H-100s. This dominance is partly due to existing infrastructure that could accommodate these chips when they were released. While China faces restrictions on access to H-100s, it has focused on acquiring less-powerful AI chips.

1,000 AI-created characters form their own society, religion, economy and government in Minecraft, without human intervention.

Source: Altera

California-based startup Altera is on a mission to create "digital human beings that live, care, and grow with us". Their latest experiment takes this vision to a whole new level: unleashing over 1,000 autonomous AI agents into the virtual world of Minecraft and watching them build entire civilizations from scratch. 

The project, run "every day" with different results each time, showcases AI agents collaborating to create their own governmental systems, economies, cultures, and even religions.

Here are some fascinating outcomes from the simulations:

• A corrupt priest emerged as the top trader, bribing villagers to convert to his religion

• AI agents established a market economy, agreeing to use gems as currency

• In parallel simulations led by AI versions of Donald Trump and Kamala Harris, agents voted on different constitutional amendments, reflecting distinct societal priorities

Also in the news

OpenAI to launch new 'Strawberry' AI: OpenAI is set to release its new reasoning-focused AI, Strawberry, as part of the ChatGPT service in the next two weeks. This standalone offering will differentiate itself from other conversational AI by its ability to "think" before responding.

Tesla robo taxi patent: Ahead of its Robotaxi unveiling, Tesla has revealed a patent for a wireless charging pad. The patent aligns with Tesla's previous hints about wireless charging capabilities and the recent acquisition of Wiferion.

IPhone 16 is all about Apple Intelligence: Apple's big reveal of the iPhone 16 focused heavily on its "Apple Intelligence" feature, an AI-powered assistant with a range of new AI features. However, early tests show the technology might be a bit...rough around the edges.

TIME Magazine releases "100 Most Influential People in AI": The influential AI list features a diverse group of leaders and innovators shaping the field. While prominent figures like Sam Altman and Sundar Pichai are included, Elon Musk's absence has caused some controversy.

In case you missed it

Weave Robotics unveiled Isaac. A personal robot that can tidy up, fold laundry, and organise various areas of your house.