![[our digital disco]](https://substackcdn.com/image/fetch/w_36,h_36,c_fill,f_webp,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35f157ee-ff8b-4484-8494-246f566b7cf7_499x499.png)
The OpenAI edition: GPT-4's red flags and black boxes.
S1E4 | Highlights and impact of this week's top tech themes.
Hi, Friends —
Welcome to [our digital disco]! I’m excited to have you here. Keep scrolling for two key OpenAI-related trends, other news highlights, and misc. outputs from my brain (Snack Time). You can also check out last week’s newsletter here.
Notable Themes
☞ GPT-4: Perks and a very, very red flag.
Last week OpenAI announced GPT-4, the latest in its series of large language models (LLMs) designed to interpret and respond to complex user prompts. The algorithm will be rolled out to millions in familiar, often subtle settings, such as LinkedIn searches and Word docs. Notably, this model can interpret and explain images, “remember” more information, and boasts 20-30% fewer hallucinations (expressing false facts with confidence). GPT-4 also significantly outperforms most state-of-the-art models — and demonstrates human-level skill — across benchmarks including calculous, law, and coding. (So far, English literature remains safe.) It can also pass written wine exams and do your taxes.
Why does it matter? GPT-4’s release and attention speaks to the AI revolution underway. Its predecessor flaunted 100M+ users only two months after its launch, a record for any consumer user app. (For reference, Instagram took 2.5 years to reach that same number.) OpenAI models are pushing the boundaries of AI research, development, and implementation. The recent explosion in other GPT-backed and GPT-inspired technologies exhibits a strong user demand for such tools, and the promise of other innovative applications.
Pros: As AI continues to advance, we can expect to see more tasks being automated or augmented by AI — leading to increased productivity, efficiency, and innovation. Case in point? Miscrosoft’s leak of Copilot, a “modern AI assistant” powered by GPT-4. Designed to streamline and automate work within the Office platform, Copilot promises tools to auto-generate PowerPoints, analyze Excel data, or summarize action items throughout virtual meetings.
The integration of AI into Microsoft's Office Suite is just one example of how AI is expediting and reshaping work. Some other notable cases: DoNotPay is using GPT-4 to sue robocallers with “one click lawsuits,” Be My Eyes uses GPT-4 to answer questions and access information for visually impaired users, and Khan Academy’s GPT-4-powered chatbot answers homework questions, suggests personalized exercises, and provides problem-solving tips.
Cons: GPT-4 poses many risks. It could be used to manipulate human behavior, provide illicit advice, perpetuate existing biases, and spread misinformation. These concerns aren’t new to OpenAI, which granted a research group access to GPT-4 to study its potential harms. The investigation — including whether GPT-4 could create and act on long-term plans, self-replicate, gather power and resources, and accomplish independent goals” — exhibits a broader concern on AI’s power over humans. The group concluded that GPT-4 was incapable of the latter goals. (To rephrase: Will AI take over the world? Not today, at least.)
GPT-4 can manipulate real humans, however.
This is where things take a turn for the wacky. The investigation did disclose that GPT-4 hired an unsuspecting TaskRabbit worker to defeat a CAPTCHA (security test to determine whether the user is human). When questioned, GPT-4 told the worker: “No, I’m not a robot. I have a vision impairment that makes it hard for me to see the images.” The person then solved the CAPTCHA for the robot.
☞ OpenAI is rebranding — by closing access to GPT-4’s underbelly.
OpenAI announced GPT-4 without disclosing information on the methods and data used to train it. The decision signified a key change from the firm’s original values: a research organization that opened its data, methods, and sourcing to the public. OpenAI explained that fierce competition, as well as concerns about who might use the information, prompted the decision. “We were wrong. Flat out, we were wrong” stated Ilya Sutskever, OpenAI cofounder and chief scientist, in reference to OpenAI’s original strategy. “If you believe, as we do, that at some point, AI — AGI — is going to be extremely, unbelievably potent, then it just does not make sense to open-source.”
Why does it matter? Sutskever’s comments address growing concerns of the risks posed by AI models as they approach human-level intelligence. (More on AGI in a future post). Yet critics argue that OpenAI's decision to restrict information about GPT-4 could hinder AI progress and safety. The decision is also part of a larger trend of OpenAI’s pivot to profit and competition. OpenAI began in 2015 as a non-profit AI research company, governed by a primary fiduciary duty to humanity, “unconstrained by a need to generate financial return.” The firm pivoted in 2019 by announcing a capped for-profit arm, and has since received investment from high-profile investors such as Microsoft ($10B).
Pros: OpenAI’s fears of competition are valid: many companies want to reproduce, and profit off, its work. Rivals are popping up everywhere to capitalize on user demand, from Elon Musk and his “anti-woke” chatbot agenda to Big Tech’s various bots to an explosion of inspired start-ups. Moreover, competitors may not robustly investigate the ethics or consequences of their products. Competition incentivizes speed over prudence — e.g., Microsoft laying off its entire responsible-AI team to prioritize deploying its products. Reduced competition might help sustain deeper research into the implications of this technology, and the creation of safeguards. Finally, the increasing availability of GPT technology raises concerns about potential misuse by groups or individuals with malicious intent (deepfakes, cybercrime). Not disclosing the full extent of the technology's capabilities could help to mitigate these risks.
Cons: Many argue that more transparency is needed to develop these safeguards and make informed decisions about GPT-4 use. Without a clear understanding of the model, users may apply it in untested or inappropriate circumstances (e.g., a rise in accidents because a self-driving car’s model wasn’t trained on appropriate weather data). Second, by keeping the research behind GPT-4 closed, critics accuse the company for going against its original ethos of “free collaboration” and instead prioritizing profit. This lack of transparency could make collaboration harder for outside researchers, concentrate power in the hands of a few large companies, and limit innovation. Centralization also raises concerns about the potential for abuse of power or the development of biased AI systems.
Other Highlights
Yesterday, the U.N. IPCC released an urgent report: We’re likely to surpass our climate target of limiting warming to 1.5 degrees Celsius by the early 2030s. Past that awaits catastrophic warming: disasters so extreme that “people will not be able to adapt.” The report pushes to expedite efforts before it’s too late.
TikTok is paying dozens of influencers to descend on Washington, DC. The influencers plan to hold a conference on Wednesday, March 22 and lobby against TikTok’s potential ban.
ALPHV, an infamous hacker group, is blackmailing Ring. ALPHV claims it has hacked the Amazon-owned camera maker, and is ready to leak internal data if the company doesn’t pay up. Ring hasn’t admitted to being hacked.
Some US senators warn that foreign governments may have helped spur the SVB collapse, and might use social media to rig the next bank run.
Snacktime
📓 Reading: Online Sleuths Untangle the Mystery of the Nord Stream Sabotage: The advantages, drawbacks, and often, false confidence generated by open-source intel.
♬ Listening to: Are We at the Beginning of a New Cold War? (The NYT Daily)
✰ Thinking about: Decisions, comfort zones, and what it means to embrace a “new normal.”
Next up
✎ Social networks and decentralization. And whatever is happening in the news. And whatever rant I go on over the next six days.
✿ As always — any and all feedback is welcome! In the meantime: give someone a hug and say an ‘I love you’ this week. Make the world a little happier.