AI Roundup 058: Devin and SIMA
March 15, 2023.
Devin and SIMA
Last year, there was a brief moment when everyone was obsessed with "agents" - AI-powered programs that could create and execute plans. Though much of the hype died down, this week saw two new agents gain attention: Devin, an AI software engineer, and SIMA, a video game-playing AI.
Why it matters:
They're still in their infancy, but capable, autonomous agents are one of the next big frontiers for AI.
With Devin, we can see what the future of software development might look like - an AI reading documentation, deploying features, and finding bugs. The programmers are understandably nervous.
DeepMind, on the other hand, has trained a versatile AI agent to follow regular instructions and perform tasks in video games, which is far more flexible than just "winning" them.
Elsewhere in flashy AI releases:
Covariant debuted the RFM-1 (Robotics Foundation Model 1), an AI model that aims to be the "ChatGPT for robots."
Figure has partnered with OpenAI to demo the Figure 01, a robot that can fully converse with people.
And Elon Musk, after repeatedly needling OpenAI, said that he plans to open-source his Grok AI this week.
OpenAI’s board gets bigger
After months of review, OpenAI has announced four additional board members: Dr. Sue Desmond-Hellmann, Nicole Seligman, Fidji Simo, and Sam Altman.
Elsewhere in OpenAI:
In addition to the new board members, an internal inquiry found that Altman's firing was due to "a breakdown in the relationship and loss of trust" from the board rather than safety concerns.
OpenAI has announced partnerships with Le Monde and Prisa Media, to bring French and Spanish news content to ChatGPT.
And in a new interview, CTO Mira Murati explains more details about Sora and how the company plans to roll it out to users.
Elsewhere in AI governance:
The EU Parliament approved the AI Act, which is expected to go into effect in May.
Some NIST staff have threatened to resign over the expected appointment of an "effective altruist" to the US AI Safety Institute.
And the US pushes for global AI regulations at the UN, in contrast to the relative lack of regulations domestically.
People are worried about AI bioweapons
In a new letter, over 150 scientists focused on protein design have signed an agreement to prevent AI from being used to create biological weapons.
Between the lines:
AI's use in creating bioweapons is a recurring theme - Anthropic raised concerns about it to senators last year, and it played a role in President Biden's AI Executive Order.
Some AI doomers also see it as a possible route to an AI "extinction event," though others argue it's extremely difficult to autonomously create bioweapons without human intervention.
The new agreement strikes a balance - calling for models to remain open while asking for steps to make acquiring physical materials more difficult.
Elsewhere in AI anxiety:
Researchers have discovered a new vulnerability in streaming LLM responses that lets attackers guess the content of chatbot responses.
Silicon Valley may be pricing academics out of AI research, as high salaries and even higher training costs grow out of reach of academic institutions.
And speaking of AI doomers, the New Yorker profiled several of them, alongside some e/accs.
Things happen
Thomson Reuters has $8B to spend on AI acquisitions to transform its business for the AI era. Palantir wins US Army contract to provide AI systems for the Army's first "AI-defined vehicle." Underdog who beat Biden in American Samoa used AI in election campaign. Meta reveals details of its new AI data centers, each with over 24,000 H100 GPUs. "Digital Marilyn," an AI avatar of Marilyn Monroe that premiered at SXSW. CBP wants to use AI to scan for fentanyl at the border. AI-generated data can poison future AI models. Midjourney can now maintain consistent characters across generations. The AI gold rush. The Truffle-1, a $1300 "inference engine" to run open-source models at home. Amazon's new AI feature creates a product page from a seller's external website. The AI-generated hell of the 2024 election. Microsoft adds GPT-4 Turbo to the free tier of Copilot, and rolls out Copilot Pro worldwide. Apple acquires Canadian AI startup DarwinAI as it tests AI-powered ads. Midjourney bans all Stability AI employees over alleged data scraping. Adobe Firefly repeats the same AI blunders as Google Gemini. Media execs say Google's AI search experience could cut organic traffic by 20-60%. The top 100 GenAI consumer apps.
