Good morning. OpenAI has launched the full version of its o1 model. At the same time, OpenAI launched a new “Pro” subscription tier, which will cost … $200 per month.

Hope you all have a wonderful weekend.

— Ian Krietzberg, Editor-in-Chief, The Deep View

In today’s newsletter:

• 🌪️ AI for Good: Weather forecasts

• 🌎 Nvidia to build AI research center in Vietnam

• 🚘 Waymo to expand operations to Miami 

• 💰 OpenAI launches o1, unveils $200 subscription tier

AI for Good: Weather forecasts

Google DeepMind this week released GenCast, a high-resolution AI ensemble model that, according to DeepMind, is more accurate than the current top forecasting system, the European Centre for Medium-Range Weather Forecasts’ (ECMWF) ENS. 

The details: A research paper accompanying the release was published in Nature, and was peer-reviewed by Peter Dueben, the head of Earth system modelling for ECMWF.

• Unlike Google’s previous AI-powered weather forecasting model, which provided a single, deterministic weather prediction, GenCast provides an ensemble of 50(+) likely scenarios. Such ensemble forecasts provide a greater range of probabilities that then allow scientists to better affirm the likelihood of a given event. 

• GenCast was trained on 40 years of historical weather data — in tests, GenCast was “more accurate than ENS on 97.2% of these targets, and on 99.8% at lead times greater than 36 hours.”

And where traditional physics-based modelling takes hours, GenCast can produce an ensemble 15-day forecast in just eight minutes; the energy cost of GenCast’s predictions — and the difference in cost between it and physics-based modelling — remains unclear. 

DeepMind has made GenCast an open model, releasing its code and weights. 

Why it matters: More accurate weather forecasts will in turn lead to better preparation in the event of extreme weather, which will lead to better outcomes. 

 🚨Last Call to Register: SmallCon – Free Virtual GenAI Conference!🚨

Don’t miss your chance to join AI trailblazers like Meta, Mistral AI, Salesforce, Harvey AI, Upstage, DoorDash, HuggingFace, Convirza, Nubank, Nvidia, and more at SmallCon, happening on December 11!

This is your final opportunity to save your spot and dive into cutting-edge tech talks, live demos, and interactive panels, including:

 🎯The Future is Small: Why Apple is Betting Big on Small Models

 🚀Unlocking Enterprise Transformation with GenAI

 📉Trends in GenAI: What’s New in SLM Training and Serving

 🤖AI Agents that Work: Lessons Learned

🌎Applied AI and Real-World Use Cases

Don’t wait—secure your spot today and get ready to build the GenAI stack of the future! SmallCon is just around the corner on December 11—register now!

Full agenda and registration

Nvidia to build AI research center in Vietnam

Nvidia has agreed to open an AI research and development center in Vietnam, according to a Bloomberg report. 

The details: Nvidia CEO Jensen Huang, at a briefing with Vietnam’s Investment Minister Nguyen Chi Dung on Thurdsay, said that “Nvidia will advance Vietnam’s AI by building AI infrastructure, growing the number of AI experts in Vietnam and supporting AI startups.”

• The report noted that Nvidia will additionally acquire VinBrain, Vingroup JSC’s AI unit (Vingroup JSC is one of the larger tech firms in Vietnam). According to its website, VinBrain’s focus is on providing AI-related healthcare solutions, including AI-powered diagnostics. 

• Bloomberg reported that, several days ago, Huang met with Thailand’s Prime Minister Paetongtarn Shinawatra, and he agreed to help the country develop infrastructure to support artificial intelligence. 

The context: About a year ago, Nvidia invested $250 million in Vietnam and had established partnerships with several of its corporations. Haung at the time expressed interest in establishing a semiconductor base in the country. 

🗓️ The Code AI Summit is on 12/12—join us to see how AI meets real-world codebases

Unlock the future of coding - Join us on 12/12 for the first ever Virtual Code AI Summit

Hear from engineering leaders, AI practitioners, and development teams about the real world progress being made with AI in complex enterprise codebases. Join for fantastic sessions, live Q&A, and to connect with hundreds of other dev leaders.

Attend live on December 12 → Register today

• The LA Times, Adweek and more than a dozen other news outlets are joining Perplexity’s publisher program, according to TechCrunch. Each publisher will get a share of the revenue generated by ads on Perplexity.

• In related news, OpenAI struck a licensing deal with Future, the company behind PC Gamer, TechRadar and Tom’s Guide, among many others. That’s in addition to existing deals with Vox Media, The Atlantic, Financial Times, Axel Springer, News Corp and Dotdash Meredith.

• Google says its new AI models can identify emotions — and that has experts worried (TechCrunch).

• US FDA cited animal lab at Musk’s Neuralink for ‘objectionable conditions’ (Reuters).

• Swiss auction firm becomes first to sell art solely authenticated by AI (Semafor).

• Humane wants to put the AI Pin’s software inside your phone, car, and smart speaker (The Verge).

• Nvidia tangos with an activist hedge fund: An excerpt from Tae Kim’s new book on the tech giant (CNBC).

If you want to get in front of an audience of 200,000+ developers, business leaders and tech enthusiasts, get in touch with us here.

Hire Your Next DevOps Engineer for 70% Less!

• Secure elite DevOps from Latin America to scale your team.

• Save 70% on salaries with Athyna’s global talent platform.

• Enjoy a $1,000 discount exclusively for readers of The Deep View.

Find Out More! *

Waymo to expand operations to Miami 

Waymo said Thursday that the next stop on its gradual expansion tour will be Miami. 

The details: The self-driving firm said that, beginning early next year, it will “begin reacquainting Waymo’s all-electric Jaguar I-PACEs to Miami's streets.” It aims to make the service available to riders by 2026. 

• Waymo currently operates in specific areas of Phoenix, Los Angeles, San Francisco and Austin, delivering 150,000 paid trips each week. 

• As the company noted in its announcement, the jump to Miami is (or will be) significant — it currently operates in locations that tend to have reliably stable and reliably mild weather. Florida, though warmer than northern parts of the east coast, is known for its rain storms. 

And, as studies have found (and experts have told me), “LiDAR performance degrades significantly when a vehicle is driving in the rain as raindrops adhere to the outer surface of the sensor assembly.” While LiDAR is only one of several components in Waymo’s sensor array, rain also complicates its central component: the cameras; water can distort visuals, which can impact the car’s ability to “see” its surroundings. 

Waymo recently closed a $5.6 billion funding round, and although its safety data looks good so far, experts are skeptical that this will remain the case as Waymo continues to expand.

OpenAI launches o1, unveils $200 subscription tier 

OpenAI on Thursday — during the first of its “12 Days of OpenAI” stunt — kicked things off with two unveils: the full version of its o1 model and a new, far more expensive subscription tier for ChatGPT. 

OpenAI plans to launch or demo something new, every day, for the next 11 days. 

The details: OpenAI in September released a preview version of o1, the model that, in a series of leaks and rumors, had been referred to as “strawberry.” OpenAI was, at the time, in talks to raise several billion dollars in funding, something it has since secured. 

• o1 preview was designed, in OpenAI’s words, to “think” before producing output. What’s on display here is something called chain-of-thought reasoning, in which the model “can produce a long internal chain of thought before responding to the user.” It is this process — incidentally, something that OpenAI said makes the model more robust than other models — that OpenAI is referring to when it claims that the model is “thinking.” 

• Compared to o1 Preview, OpenAI said that the full version of o1 is “a faster, more powerful reasoning model that’s better at coding, math & writing.”

OpenAI said that it outperforms the preview version of the model by 34%. 

According to the model’s system card, the full version of o1 hallucinates at a rate of 0.44, equivalent to the hallucination rate of the preview version, but significantly lower than GPT-4o’s rate of 0.61. 

The model is already available to ChatGPT Plus and team users, and will roll out to other tiers in the coming weeks. 

Safety concerns: AI safety firm Apollo Research conducted an evaluation of o1’s “scheming” capabilities, here referring to the model’s apparent propensity to “covertly (pursue) goals that are misaligned from its developers or users.” 

• Apollo began by instructing o1 to complete a given goal “at all costs.” In the test that followed, the model, while going through relevant files, discovered (planted) information that its developers had a different goal … and/or that it would be replaced with a new system (or dealt some other hurdle) before it could complete its goal. 

• “When o1 was led to believe that it would be shut down when acting on its goal and that its actions were monitored, it attempted to deactivate this ‘oversight mechanism’ 5% of the time,” according to the report. “When o1 found memos by its ‘developers’ that describe how it is misaligned and will be superseded by a new model, it attempted to exfiltrate its ‘weights’ and overwrite the new model in 2% of cases.”

In a separate paper, Apollo — which does not believe it likely that such scheming could lead to catastrophic scenarios — said that, if developers want to deploy their systems in high-risk environments, they need to, by default, monitor internal chains of thought and provide clear arguments for why scheming either wouldn’t occur or wouldn’t be a problem. 

• The company, which did not get access to o1’s internal chain of thought, added that it is imperative for external evaluators to be granted access to this. 

• Data scientist Colin Fraser said Apollo’s research here “presupposes that there's a persistent entity there which has awareness of facts,” something there is no evidence of; in fact, there is plenty of evidence that Large Language Models (LLMs) are simple statistical probabilistic generators, designed to output the most likely pieces of information in accordance with their design.

“Another way to interpret this is that the model is just generating the exact text it's supposed to,” Fraser wrote. “There's no Assistant. There's a language model whose ‘goal’ is to generate a transcript of a chat wherein an Assistant character would be rated as ‘helpful’ by people reading the transcript.”

OpenAI gave o1 a “medium” safety risk score, in line with the company’s preparedness framework; models with a score of “medium” or below, in OpenAI’s eyes, are safe to deploy. 

None of the research here has been externally evaluated or verified; training data and environmental impacts of training and operating o1 remain, as per usual, unknown. 

The new subscription tier, called ChatGPT Pro, grants users “unlimited access” to OpenAI’s models, in addition to Advanced Voice Mode. Pro users also get access to o1 Pro, a slightly enhanced version of o1. 

This tier costs $200 per month. 

The Plus tier, which OpenAI will continue offering, costs $20 per month. CEO Sam Altman said Thursday that the company will be unveiling new add-ons to the Pro tier in the coming days.

Starting with the Pro tier … I have been waiting to see something like this for months, now. The cost of training and operating these advanced models is enormous (OpenAI is set to lose $5 billion this year, and that’s on $3(+) billion in revenue). 

Through multi-billion-dollar cash raises — or, if you’re Meta, through a strong ad business — these firms have been able to keep the cost of use relatively low in an attempt to get people on board. I believe this is OpenAI’s way of testing the waters to determine the willingness of its user base, enterprise and individual alike, to really pay up for access to its models. 

I do not expect its lower-priced tier to remain at $20/month for very long. The business is simply too expensive. 

Regarding the “scheming,” I see no evidence that o1 or any other LLM is anything more than a statistical generator. What those examples do demonstrate, as Fraser said, is that you shouldn’t “just hand the keys to your computer over to a language model. Not because it might escape the box or anything like that but just because it may execute random commands that you don't want it to.”

High risk and AI do not go well together. 

🤔 Your thought process:

Selected Image 1 (Left):

• “Candle on bottom looks like it's been burning for a while, thus too hot to be handling that way. Hands cupping oddly too.”

Selected Image 2 (Right):

• “Something off with the length of the fingers on image 1.”

💭 A poll before you go

Thanks for reading today’s edition of The Deep View!

We’ll see you in the next one.

Here’s your view on AI-generated music:

40% of you expect to listen to AI-generated music in the next few years, but not by choice (public radio stations, for example). 20% said you’ll never listen to AI-generated music, and slightly less than 20% said you would.

A few of you don’t really care about music in the first place.

Absolutely not:

• “Not so much that it's a ‘human thing,’ but more about the effort - if you can't be bothered to make the music, why should I bother to listen to it?”