Trusted #012: Wrapping 2023 and Previewing 2024
Fast-tracking education to keep pace with the world
Author’s Note: Where did I go for the last three months? That requires some explanation.
At the beginning of 2023, I realized that AI/LLM’s were going to be an inflection point for technology, and I wanted in. Unfortunately, the military’s not one for rapid career pivots. Making the best of it, I decided to set myself an aggressive educational timeline to learn as much as I can.
In short: With no formal CS background1, I started AND completed an undergraduate degree in computer science from WGU and will be starting a master’s degree in computer science at Georgia Tech next year, while continuing my full-time military duties. I’ve essentially done nothing but work/family/classwork for the final four months of 2023.
I may write up a more detailed guide if there’s interest, and I’m already second-guessing it a bit (if I can do something that quickly, exactly how much value am I gaining from it?). Still, it was more productive than many of the other things I was doing with my free time. Now that I’m off the “speedrun college” track, I should be able to write more often. (Famous last words.)
Well, that was a year, huh?
Since ChatGPT came out in late 2022, the expectations in the tech landscape have been reshaped dramatically. Now, I’m choosing my words carefully here: The landscape itself isn’t too different, except in a few key places, but people’s projections for how the future will play out have changed dramatically, which is driving investments and other decisions.
As I’m getting my writing muscles warmed back up for 2024, here’s my roundup of the most impactful stories from last year.2
OpenAI continues raising the bar: OpenAI is not resting on their laurels. After deploying GPT 3.5 Turbo/ChatGPT (the “fastest-growing consumer application in history”) last November, they released GPT-4 midyear, which remains the highest-quality model by a significant margin. They added “Code Interpreter” as a separate mode (now incorporated automatically) for data analysis and programming support in July, and added image generation via DALL-E to ChatGPT later in the year. It hasn’t all been great; the attempts to build an “app-store” like model, by allowing external plugins and later “GPTs”, haven’t amounted to much so far, and there was that whole “CEO gets fired and rehired a week later” drama that totally consumed everyone’s attention over Thanksgiving.
Microsoft and other Big Tech companies come to play: In January, Microsoft announced a $10 billion investment into OpenAI. It quickly followed that up with a rollout of “Bing Chat”, essentially a slightly tuned ChatGPT, in February, which led to a rapid retraction after the original chatbot was…unprofessional. Despite that, Microsoft has made AI a centerpiece of their strategy, jamming “copilots” into everything. Google has followed suit in a more measured fashion; Google rolled out the widely-panned Bard in March, and updated it via the new “Gemini” algorithm a few weeks ago (still not that great, pending rollout of the full Gemini model next year). Meta is going a different direction, releasing LLaMA in February and LLaMA 2 in July as “open-source” models. Anthropic (supported by Amazon) released Claude/Claude 2, which are the current best alternative to ChatGPT.
AI Safety advocates shape the discourse: Accompanying the AI hype has been calls for safety and regulation, with no clear consensus on the necessary measures. (My very first post broke down views about AI Safety into four groups, and is still worth reading.) The Future of Life foundation posted an open letter in February calling for a “six-month pause” that received significant attention. Sam Altman and Gary Marcus spoke to Congress about it. A second “open letter” from the Center for AI Safety was boiled down to a single sentence, and received much more support in June. Finally, two regulatory measures near the end of the year came out: the US executive order on AI, and the EU AI Act, though a significant amount of time is needed to determine what effect, if any, these efforts have.
2023 Product Recommendations
For posterity’s sake, I thought it would be good to list out some of the products I’m using currently, as I expect this to change significantly by next year.
Best consumer LLM, paid: GPT-4, accessed through OpenAI ChatGPT Plus ($20/month). Pay the $20. GPT-4 is still a significant quality increase to anything else that is commercially available as of today. I use it on a daily basis.
Best consumer LLM, free: Microsoft Bing Chat, Creative Mode. Bing Chat (soon to be renamed Microsoft Copilot) in Creative Mode uses GPT-4 under the hood, and is free; the downside is the Microsoft flavoring on top makes it a less reliable option, and you lose a lot of the OpenAI-specific features. Honorable mention: Anthropic’s Claude (good for documents and probably second best model overall), and , and of course the original ChatGPT. Inflection AI’s Pi (for grandparents) or Character.ai (for Gen Z) are also interesting but wildly different chat-focused experiences. Poe AI is good for experimenting with different models, as it has APIs to most. Perplexity AI has really good search features. Google Bard, after the Gemini Pro update, is okay but not as impressive as claimed.
Best consumer LLM, opensource/edge: Mistral AI’s Mixtral. Coming out of nowhere just a few weeks ago, this French model edges out Meta’s LLaMA 2 in quality, though it’s brand new so still rough around the edges. Getting it up and running locally is beyond the scope of this guide, but Poe (linked above) offers it as an option if you want to experiment.
Best image generation model, paid: Midjourney. I haven’t used image gen as much recently so I don’t feel as qualified to comment, but others I trust say that Midjourney is still the best. I use DALL-E 3, which comes with ChatGPT Plus and works for my limited use case.
Best image generation model, free: Bing Image Creator, more for its wide accessibility than anything else. Built on DALL-E 3.
Honorable mention: If you’re in the US and have a Meta account, Meta’s Imagine is good. Adobe Firefly has a daily free trial, as does Ideogram.
Best image generation model, opensource/edge: Stable Diffusion XL. There are several websites offering various versions of Stable Diffusion, though be extremely careful as Stable Diffusion is commonly used for “uncensored” imagery. Playground AI seems relatively safe.
Predictions for 2024
And finally, purely for fun, here’s some 2024 predictions.
A substantially AI-generated song charts in the Billboard Hot 100 at some point during the year. “Now and Then” opened the doors for this, and I think you’ll see at least one label try to generate something new using the vocal style of an artist that has passed away. I would be surprised (but not shocked) to see a VTuber/Hatsune Miku-type of thing go viral on TikTok and sneak in, though.
Voiced AI Characters have a moment. I’m not sure how exactly; maybe a celebrity leaves his/her partner for a Replika, maybe Meta decides to create “virtual Facebook friends”, I don’t know. By the end of 2024, it won’t be normalized, but it won’t be new, either.
Limited-domain AI Agents start rolling out; most are ineffective, but there are a few standouts. Google is just testing this now, for example, with connecting Bard to Gmail and Gdocs; by the end of 2024, I think you’ll see this everywhere, and it’ll be mostly ignored because it sucks, but a couple places will pull it off.
ChatGPT remains the gold standard for LLM’s, but the gap shrinks. I expect Google Ultra3, when it gets rolled out in January/February, to approach GPT-4 level performance, but I will be surprised if it passes it. I then expect a midyear GPT-4.5 style update that will open the gap again, and so on.
AI Safety will be a talking point during the US 2024 election cycle. Not a permanent talking point, but it will come up, get discussed, and then dropped again.
I’m glad to be back and writing again, and excited for my first master’s level course to destroy my current optimism! I’ll see you in 2024!
Standard disclaimer: All views presented are those of the author and do not represent the views of the U.S. government or any of its components.
I do have a significant IT/cybersecurity background, which I received some credit for, and I did get general education credit for my previous bachelor’s degree. It was two years worth of schoolwork in four months, not four.
All news links to the New York Times. No personal endorsement implied.
Sadly not Google Ultron.
Congratulations for the CS degree ! Always learning is a great attitude and achievement