Breaking Down the Best Language Models of the Year
And what they each excel at
We’re about to finish the first entire year with access to AI for us muggles and we’ve now got incredible tools and assistants for pretty much anything.
And especially learning.
I’ve followed AI advancements throughout the year and kept an even closer eye for the past 6 months. I’ve also experimented tons, mostly with ChatGPT but also the other major Large Language Models (LLMs) like Bard or Claude 2. I’ve also tried out some others worth mentioning but let’s not get ahead of ourselves.
I’ll compare the main free AI chatbots but will also mention the paid GPT-4 to indicate the current best on the market.
Remember though, I’m no tech expert so I’ll try to keep things quite simple and more on the what-can-we-do-with-this level.
Also remember everything changes often in this world, as you’ll read on. For this reason, I’ll update this page every few months.
First, here’s a table about the basic stuff.
What AI chatbot for what?
As you can see, some basic facts can help us choose the right free tool already.
For example, if we need to check online, Bard is the only option. If we want to check some info in some files, that’ll be Claude 2. And if we want to have the chatbot follow strict rules, ChatGPT’s custom instructions are our best bet.
Of course, GPT-4 allows all of this and more, but that’s paid so that’s not fair.
Let’s dig a bit deeper into each tool’s clear advantage.
Bard’s top points
Google’s Bard recently got an update, 20 days ago, with Gemini Pro added to it. What it did was theoretically get better than ChatGPT 3.5 when it comes to results.
My own tests showed ChatGPT was still better when given specific directives but, then again, I don’t use ChatGPT 3.5 unless I’m experimenting for this newsletter so I originally thought I just got lucky. But it now seems confirmed: ChatGPT is still better.
Where Bard shines is its connection to the internet. And I’m not just talking webpages. I’m talking YouTube videos too.
You see, Bard connects to YouTube, Google Docs, etc.
As a learner, this is a godsend.
We can ask Bard to give a summary of videos before we watch them so we start with an overall idea of the topics. Or we can turn to it after watching the video to check our understanding and even dive deeper into one of the topics.
Now, let’s not get too excited though. For language learners, this doesn’t seem to work for anything other than English right now. It was announced to work in Japanese and Korean on December 18th but I haven’t gotten it to work with a single video in these languages yet.
I’ll keep an eye on it and will let you know when it works for other languages!
Still, a video you loved in English could work as a trigger to learn vocabulary about that topic in your target language.
Another Bard-specific advantage is the capacity to upload images. While this may not seem so useful at first glance, this also means we can work through those and develop our understanding of its content.
As a language learner, this means I can develop my capacity to describe the world. For example, if I upload a picture of Korean mountains and ask for an explanation, I can get the following explanation:
Unfortunately, this doesn’t work with pictures of humans so it’s not helpful to learn to describe people or expressions.1
Some other advantages of Bard include:
Getting up to three drafts of responses
Can get an audio version of any answer, with a quite-alright voice.
Proper cons are:
No app so not convenient from phone
Inconsistent for long conversations (doesn’t follow directives well)
Only possible to edit the last prompt so impossible to create two forks of conversations and see where each will take us.
Claude 2’s top points
Now let’s turn to Claude 2. This LLM is much less talked about outside of AI-curious circles.
For the short story, Claude 2 was created by Anthropic, a company created by some ex-employees from OpenAI who wanted to focus on the research aspects. While not at the level of ChatGPT or Bard, it’s still a pretty good chatbot.
And it has two incredible advantages compared to other tools:
Allowing uploads of files. It accepts PDFs, Docx, csv, txt, and Html.2 You might have noticed I used the plural. That’s because we can upload up to 5 files per conversation.
Understanding up to 200,000 tokens (~150,000 words), which means—in theory—it can pick up information anywhere in that large of a window. Unfortunately, this suffers all the more from the lost-in-the-middle phenomenon I mentioned in a previous piece.
For today’s students, this means Claude can refer to PDFs received by the teacher or tutor and help you work through them. It can create exercises to master some parts or help you work through difficult points with more explanations or examples.
When it comes to cons, I’d say Claude 2 is just average for most things as of the end of December 2023. So unless you need to work through multiple or long files, you might as well use other tools.
ChatGPT (3.5)’s top points
ChatGPT doesn’t need much of an introduction anymore, does it? This tool created by OpenAI was the trigger for AI to spring all around us since November 2022.
ChatGPT has been great at what it does since the beginning but it got an update in November making it much better (again, theoretically).
ChatGPT 3.5’s biggest selling point to me right now are as follows:
It’s been written about tons online so tips and hacks to use it better are everywhere.
You can use Custom Instructions to which ChatGPT will refer to before each answer. This allows for much better long conversations following set rules. I’ve written about it more here.
It has an app, making it convenient to use from anywhere!
The app has a Voice function to speak directly to ChatGPT in real time. There also isn’t much waiting for an answer. It works well and each of the 5 voices is quite good.
I’ve spoken tons about ChatGPT so I won’t ramble too much here.
It’s a versatile tool I think everybody should use daily. With the right set of custom instructions and a good knowledge of how to prompt it well, it can help with pretty much any knowledge task.
Learning included of course!
Bing Chat’s top points
Now, let’s turn to one I had kinda forgotten until a recent newsletter from
.Microsoft partnered with OpenAI some time ago and made it so Bing Chat could access GPT-4. Yes, GPT-4, the paid version. Don’t ask me how that makes sense though.3
What does this mean?
It means Bing Chat (also called Copilot now) can:
Access the web, through the “incredible” (🤭) search tool Bing and study through what you find like you could with Bard.
Create images with DALL-E 3 here or in the Chat itself.
But that’s not it, it also has three settings we can use:
“More balanced” → best for “most common tasks” and fastest mode
“More precise” → most factual
“More creative” → more expressive, longer output but slowest mode
One aspect I quite liked was that the fact Bing proposes three possible answers we could use to reply to it, which is great when you want to keep the conversation going but don’t know what to say anymore.
This being said, it’s important to note the maximum number of responses we can get is 5 per conversation (or 30 if on the app). We can create up to 15 conversations per day. This makes Bing Chat not a great tool for practicing proper conversation skills in a new language.
Now, this leaves out where I think Bing Chat shines, something I mentioned two weeks ago, discovered in this piece from Lou.
Bing Chat has a voice function to hold spoken conversations. This has the same conversation limit but does one thing even better than the paid GPT-4: it’s possible to read the audio answers as we’re listening to them.
Unfortunately, no choice of voice here.
Unfortunately, the voice function doesn’t automatically recognize the language. Instead, you have to set the app’s language to the language you will record in.
GPT-4
And this brings us to the GPT-4, the paid version of ChatGPT. In short, it does everything mentioned above except two things:
Upload multiple files (only one per message, although it’s possible to have many per conversation)
Voice conversations on the app are voice-only. It’s possible to see the text (transcript of the conversation) when closing the voice feature though.
It also had many more advantages such as:
It’s on a whole other level when it comes to following prompts.
It’s possible to create “custom GPTs” with a specific set of rules for precise uses, such as practicing specific aspects of a language.
You can also use custom GPTs for other things too. For example, I can’t seem to remember the conversion between kilograms and pounds. My gym has weights indicated in lbs so I always need to check on my phone. I’ve set a GPT to respond to “X lb” with just “Y kg” and no other text. Could I use Google for this? Sure, but this is faster to access on my phone (and it avoids opening yet another tab.)
Recommendation for language learners
Before we move on to some bonus tools worth keeping an eye for, here are my recommendations based on the type of skill you want to practice in your target language.
Discovering new grammar patterns: Bard (to get links to websites/videos)
Practicing grammar patterns: ChatGPT (with Custom Instructions for a “teaching” style)
Reading recognition practice: Bard (to listen to multiple draft answers to the same prompt)
Reading comprehension with real texts: Bard (asking for a summary of content in a link)
Reading comprehension: ChatGPT (asking for a story to be created and working through it.
Written conversational skills: ChatGPT
Oral conversational skills (when starting or to practice difficult topics): Bing Chat
Oral conversational skills (without written support): ChatGPT
Of course, to each their preferences.
If you have the time, I recommend experimenting and finding what you prefer!
Other tools worth considering
Now, I know I’ve focused on the main tools but there are others worth knowing about and even trying. Some might become contenders in 2024.
Pi.ai has been around since May 2023 but I only heard of it in late November when the announcement of Inflection-2 came to indicate it could be better than Bard (at the time). It’s still not out yet though.
Pi.ai currently only understands English so it won’t be helpful for language learners who already know English. It can be useful for other learners but it seems to be shining as a mental support for those in need.
Then, there’s Perplexity.ai, I used to use it to connect to the web before Bard came out but there’s not much of a point anymore for this. Still, it has two aspects worth mentioning:
It has an app which makes it better than Bard in this sense
It has a browser extension which is great for getting summaries of pages you’re on without leaving them. It even works in other languages.
Finally, there’s Character.ai. This is just a conversation partner tool that takes the character of a specific person. For example, you can talk with a personification of Socrates. This is something ChatGPT does well too but I guess the UI is cleaner and it can help separate topics too.
I thought Character.ai’s big advantage for language learners would be the possibility to learn to imitate the style of speaking of a specific person but the problem is it only seems to work well in English. When I spoke to “Socrates” in French, a grammar mistake happened in the second answer!
Also, the “Hyperglot” character doesn’t have a specific style as far as I’ve seen.
Final thoughts
There’s a lot out there and there’ll be even more by the end of 2024.
Still, I recommend trying them all out and finding which you like for which purpose. As I’ve mentioned before, there’s no turning back. AI is here to stay. I read somewhere it was estimated about 80% of the population would use AI daily by the end of 2030.4
This seems high but I reckon we’ll reach this percentage for people under 40 before the end of 2025 though.
Does this mean we don’t need to learn anything else than AI anymore?
No, but it does mean we should use it to help us learn.
As language learners, we now have the tools to create our own tools for dual-coding. Pretty much anything we want can now exist both in audio and text format. We can even have images along. I also didn’t mention videos but “text to video” AI tools are also getting better with tools like Pika.ai.
2024 is gonna be a great year of discoveries. I look forward to it and hope you’ll join me for the ride!
Cheers for reading,
Mathias, a forever student of things
PS: Piece based on tools as of 12/26/2023
I tried jailbreaking it (ie. working around the limits to push it to do something it’s not supposed to do) without much success.
Unfortunately, no Excel yet but that’s less important for language learners.
The few tests I’ve done still gave better results with the “real” GPT-4 I paid for.
I can’t seem to find the news article anymore unfortunately but either way, it’ll be high.