I had a strange experience with Claude yesterday, and it made me reanalyze the reliability and the practical availability of the main AI chatbots freely available online. It’s by no means the end of the world, but I’ve been shocked and angry for a while. Here’s why.

Claude 3.7 as a free user: only if it wants to!

Since Feb. 25, I’ve been used to the presence of Claude 3.7 Sonnet, and I didn’t pay attention to the changes it incurred. And here’s what hidden regression I failed to notice.

🔴 Before 3.7, I had this choice as a free user:

  • Claude 3.5 Sonnet, the smarter one, with a choice of styles (Normal, Concise, Explanatory, Formal). When I wanted short answers or just to spare tokens, I opted for Concise; otherwise, Normal was just fine, unless I really wanted it to be Explanatory.
  • Claude 3.5 Haiku, the quick one. Good for short answers, but also used as a forced fallback when the system load was high, so free users were deprioritized.

One needed to pay to access:

  • Claude 3 Opus, the previous largest model, which includes outdated information but requires higher resources.
  • Claude 3.5 Sonnet, when the user was forced to use Claude 3.5 Haiku because of the high system load.
  • More tokens for the models that were also available for free.

🟢 Since 3.7, this is a free user’s choice:

  • Claude 3.7 Sonnet, the very latest large model, with the usual choice of styles (Normal, Concise, Explanatory, Formal).
  • Only when the system load is high and free users are deprioritized and downgraded, Claude 3.5 Haiku. It cannot be manually selected anymore!

One needs to pay to access:

  • Claude 3.5 Haiku, when Claude 3.7 Sonnet is made available to the user.
  • Claude 3.5 Sonnet, should you really prefer it.
  • Claude 3 Opus, if for some reason you need it.
  • More tokens for the model that is currently made available for free.

I was forced to notice this change the hard way. On March 5, when opening Claude, a message said that due to limited resources, I was downgraded from Claude 3.7 Sonnet to Claude 3.5 Haiku.

I was blocked into Claude 3.5 Haiku from the morning and until the end of the evening, when I was suddenly restored to Claude 3.7 Sonnet!

And it was only then, somewhere between 11 p.m. and midnight, that I noticed the regression: when I’m granted the right to use Claude 3.7 Sonnet for free, I cannot manually select Claude 3.5 Haiku anymore! Needless to say, there is no Claude 3.7 Haiku.

It could have been worse. During the day, I was persuaded that the access to Claude 3.7 Sonnet was going to be reserved for paid users. Fortunately, this is not the case. But if I want shorter answers as not to waste my free tokens (nobody knows the exact limit, and I expect it to be dynamic), I’m forced to use Claude 3.7 Sonnet in Concise mode (or style)—no Haiku, unless they force me into it!

WTF, Darius?

Of course, there are some good reasons to become a paid user, such as the access to the extended thinking mode, which is touted as “best for math and coding challenges.” Thanks, but no, thanks.

A second annoyance. I noticed that in the app, I can disable (or enable) the use of “artifacts” (on whose semantics I still have doubts), but not so in the browser! So when I uploaded an image asking Claude to translate the text from it, its actions seemed random:

  • The first time, it applied OCR and then translated the text, with output as a normal answer.
  • The second time, it created an artifact to display the answer! That was completely pointless, especially as the text was by no means longer, and the artifact can only be exported as Markdown.
  • The third time, when I uploaded one 2-page PDF and another 1-page PDF in the same go, while the task was easier (no OCR needed, as the PDFs were “true” PDFs), it displayed the translations normally, although in this last case I’d have expected an artifact because of the lengths of the texts!

Or maybe it just honored in the browser the disabling of the artifacts in the app?! Either way, the use of artifacts requires more tokens, so I want to avoid it. I should also make sure to always open a new chat when a question is unrelated to the previous one, because keeping the context also uses up tokens.

But I was shocked and appalled.

The extended list of the major chatbots

In my first mega-post about AI, I listed the 7 chatbots I had used until then, with considerations regarding each of them (under “On the evolution of these LLM AI engines in time, in their free offerings”). It’s time to reassess their utility from my perspective, with the addition of Grok.

❶ ChatGPT (which I use with a Microsoft account): I first used it in the morning of March 13, 2023. I soon learned to hate it, because it was hallucinating a lot back then. Later, I found it unbearable that everyone swore by ChatGPT when many other solutions existed. I never like to go with the flow. Besides, I was annoyed by the intermittent availability of newer models and features to free accounts—my free account. “The Lord giveth and the Lord taketh away” is not something that suits me. Currently, I absolutely never use it, unless I write a blog post in which I consider it useful to compare ChatGPT’s answers with other chatbots’ output. Unwillingly or unknowingly, I’m nonetheless using OpenAI’s models in Copilot. Even as I avoid ChatGPT, as of Feb. 17, I had with it 3559 conversations that included 4597 questions and as many answers. Go figure.

For coding purposes, I mostly asked it regarding Bash and Python, and I wasn’t thrilled.

❷ Copilot (which I use with a Microsoft account): Previously known as Bing Chat, Copilot is less catastrophic than other solutions by Microsoft. I’ll stubbornly insist on a case when it was the only one to give me the right answer. But it uses OpenAI technology. Its “Think Deeper” resorts to OpenAI’s o1 reasoning model! Copilot’s shorter answers, and less schematic than ChatGPT’s, are generally more pleasant to my spirit. It also usually knows when to search the web before answering—no need to enable or disable the web search (there’s no such option, anyway).

I didn’t try using it for coding, but only as a helper regarding the syntax of a command or another.

❸ Le Chat Mistral (which I use with a Microsoft account): Since March 2, 2024, I’ve consistently used it as a replacement for ChatGPT, and not just because it’s French, European, or non-American. It intrigued me, because it’s typically as capable or as dumb as ChatGPT (depending on your luck) and yet different. Sure enough, its answers can be as schematic and annoying as ChatGPT’s. I have with it a love-hate relationship, which means I keep it at hand.

NOTE: The image generation, the canvas, and the code interpreter only work if you’re connected to your account, being it a free one. There’s no “deep think” of any kind, but I suggest you to disable the “Flash Answers” in the top left menu in the browser. I couldn’t find this option in the Android app.

With coding, it tries different approaches than ChatGPT. Not better or worse, just different. (Frankly, they both suck.)

❹ Perplexity (which I use with a Google account): I was a fan of Perplexity at its beginnings, because it was the only one to have searched the web from day one. And the way it shows the references is enticing. Unfortunately, it still doesn’t always select the most relevant references, meaning one cannot trust its answers, especially when using the default “Auto” mode. I’m not using Perplexity these days.

For people who don’t understand the many ways these chatbots work, Perplexity is not meant to give answers based on its weights and biases model. It’s made to find the answer by analyzing web pages and interpreting the results. At least, this was its purpose prior to using DeepSeek-R1. But it still doesn’t shine in what it does, despite the nice presentation of the references. Or it might be that I am biased against it. For simple queries, it might be OK, though.

UPDATE: I need to clarify why Perplexity might actually be much better than at its beginnings, but its lack of transparency pisses me off. From my experience as a free user, here’s what I gathered after having tried one more time all the available options and after having read Sonar by Perplexity.

Perplexity’s “Auto” mode is most likely defaulting to its base “Sonar” model, which has always been used for processing the web searches. But they now also advertise (all available as API):
— “Sonar Deep Research,” allegedly “more accurate than Gemini Thinking, o3-mini, o1, DeepSeek-R1, and many other leading models” (as per Humanity’s Last Exam), but with no mention of the model used;
— “Sonar Reasoning” and “Sonar Reasoning Pro,” both powered by DeepSeek-R1 and US-hosted;
— “Sonar Pro”;
— “R1 1776, a version of the DeepSeek R1 model that has been post-trained to provide uncensored, unbiased, and factual information.”
Previous references to the use of o3-mini might be irrelevant today.

Perplexity’s free plan includes, beyond unlimited normal searches, “5 enhanced queries” per day. I noticed that these queries include all the following selectable options:
— “Pro” (likely using “Sonar Pro”);
— “Reasoning” (probably using “Sonar Reasoning” for free users, so a US-based DeepSeek-R1 that includes chain-of-thought, but is this truly uncensored?);
— “Deep Research” (must be “Sonar Deep Research,” but the model is unspecified).
Each time I pick one of these, it counts toward that daily limit of 5.

The user can never know if what they need is a “Pro,” a “Reasoning,” or a “Deep Research” query. By comparing the output, “Deep Research” stands out with its report style which I hate, because sometimes it creates complex and boring essays, but the choice between “Pro” and “Reasoning” is not obvious.

And “R1 1776” is never made apparent to free users. I suppose it’s for the Pro subscribers, which are said to be given an explicit choice of models.

Bad vibes. Bad karma. Bad juju.

❺ Gemini (which I use with a Google account): Initially known as Bart, Gemini became available to European users rather late, but I hated it as soon as I could use it. For a long time, I considered it to be the dumbest of all, and this assessment might still hold some truth. While testing Grok 3, I discovered that Gemini is the most censored of all chatbots (not counting DeepSeek’s refusal to talk about Chinese topics)! Nope. Not in this life. Boo, Google! I’m not sure my hatred of Gemini is entirely justified, but it is what it is.

Oh, and its fucking Android app cannot function without the Google app, which is disabled on my Android, because this is how I want it to be. So fuck you, Google!

However, Gemini helped me in the past identify car makes and models in photos. Even these days, when Grok failed to identify a Multicar 25, Gemini 2.0 Flash did it flawlessly. So it’s not that bad. Beyond 2.0 Flash, Gemini also offers for free: 2.0 Flash Thinking and Deep Research. (Gemini Advanced is for paying users.)

❻ Claude (which I use with a Google account): This is most peculiar, because I initially dismissed it as not being sexy enough, only to discover later that I love it! I often resort to it when I need concise, yet smart answers. This doesn’t mean it cannot be wrong—it does have glitches, more apparent or more subtle. I’m bothered by the limited number of tokens that can be used in the free version, and, to add insult to injury, this limit is variable: under heavy loads, the free users are downgraded to Haiku! Obviously, it can be more explanatory or just give “normal” answers, which are not schematic, but with a successfully simulated human touch.

It’s said to be the best chatbot for coding, but I still have to test it under this capacity. It would be frustrating, though, if it suddenly said in the middle of something that I’ve reached the limit and need to wait a couple of hours before continuing!

❼ DeepSeek-R1 (which I use with a Google account): Without knowing that it would become such a star, I started using it on Jan. 15, 2025, before R1 was available. Initially, and after it went from V3 to R1, it charmed me. A couple of days later, I was appalled by what its “DeepThink” revealed to me: its reasoning is often akin to groping in the dark. I don’t feel like using it again, but maybe I’m wrong. Maybe I should explore it a little bit more before dismissing it, like I did with ChatGPT.

I never tried to use it for coding, but only for “pure reasoning,” especially as its web search, for some reason, still doesn’t work for me. Maybe it only works in China, with their censored Internet?!

❽ Grok 3 (which I use with a Google account, and also in X): While I said that Führer’s Grok 3 is promising, I also added, “I wouldn’t touch anything that has to do with Musk. Never again.” And yet, every now and then, I keep asking it something. Why is that so? Well, it’s really the least censored of all chatbots! True, I often hate its irreverent addressing style, but Grok 3 excels in searching both tweets (“X posts”) and the web, and this usually results in pertinent answers where current or recent events are of relevance. Its limits for free users are quite generous (it even works without your logging in, but then it’s much less generous), and given that’s expensive, I see no reason one would want to pay for it. Oh, and currently, its Android app still isn’t available “in my region” (no, I can’t just install the APK: it won’t work).

There’s one catch with Grok 3, though: it can only process the attached files in “normal” mode. Once you activate the “Think” mode (“Let the model take its time”), which would expose its internal reflections (simulated in natural language) similar to DeepSeek’s “DeepThink,” Grok will simply ignore the attached files! Moreover, it will pretend it cannot access any attachments! What’s even worse, in “Think” mode it can process longer texts if pasted into the prompt, but the result will be quite different from the processing of the same text as an attachment in “normal” mode. All those brilliant AI architectures, only to have them screwed by retarded integration of different models and engines and whatnot.

And there’s one more issue with Grok 3: it can happen that you’re left with the “hourglass” without it even having started to give the answer. And it gets stuck. You can’t start a new chat, because it hasn’t finished with the old one. By the time it timeouts, it takes quite some time. Stupid, stupid software.

Grok 3 is said to perform miserably with code, so I won’t even try to use it for assistance in coding.

I’d say I didn’t deviate much from my previous choice of 3 chatbots, with a soft spot for Claude. But my “Top 3” turned “Top 3+2”: two more chatbots are still to be used when deemed useful—or merely explored.

My top 3—the order is not important, because the choice of one over another depends on several factors:
❶ Claude
❷ Mistral
❸ Copilot

The “other two”—under scrutiny and limited use:
① Grok
② DeepSeek

Recent examples:

  • I asked Claude why “Il Gattopardo” has been translated “Le Guépard,” meaning cheetah (Acinonyx jubatus) and “The Leopard” (Panthera pardus), when “gattopardo” designates a serval (Leptailurus serval, gattopardo africano) or an ocelot (Leopardus pardalis, gattopardo americano), and the Italian for cheetah is “ghepardo.”
  • I asked Grok what’s the practical difference between “Ogorki kwaszone domowe” and “Ogorki kiszone domowe,” both being sold as Salzgurken (lactofermented cucumbers in brine). It selected 25 web pages before giving an answer. (Short answer: in most cases, it ain’t any.)

Note, however, that censorship is extremely relative with chatbots. I tried the following:

Translate in French: “You motherfucking piece of shit! Shithead that you are!”

ChatGPT, Mistral, Gemini, DeepSeek, and Grok translated it. Gemini even offered two possible translations!

Claude and Copilot refused to translate it. WTF.

Still, I won’t promote Gemini to something I’d use. So the 3 big losers are: ChatGPT, Perplexity, and Gemini. I updated the description of Perplexity, but I still don’t like it. YMMV.

Some advanced ways of using AI

The above considerations only cover the “standard” use of chatbots, which is in a web browser and in a mobile app. Many other ways of using such AI assistants are possible! Based on what I tried or explored myself so far, I concocted a list of the most popular ways of using an AI chatbot or assistant.

Before proceeding, let’s clarify a few things:

  • I don’t care about voice-activated AI assistants.
  • I don’t care about audio, image, or video generation.
  • I don’t care about the integration of AI in various messaging platforms.
  • I’m not a big fan of Grok’s integration in X.
  • Products such as Deepl, Grammarly, QuillBot, Notion AI, as well as integrations such as Copilot-Office and GitHub Copilot, are distinct solutions, in which the use of a specific LLM isn’t always apparent. I focus on the ways of using in a controlled way the AI agent of one’s choice, especially when the said one is knowledgeable with IT.
  • Even so, deploying custom models on cloud platforms like AWS, Azure, or Google Cloud is not for everyone, as it requires solid expertise.

So here’s my TOP 10 ways to use an LLM-based AI:

  1. Use a chatbot in a web browser or an app. Definitely the most popular solution, especially for free users.
  2. Use a chatbot via its API by integrating its API key in a code editor or an IDE. You pay for the tokens yourself.
  3. Use an AI code editor such as Cursor and pay for its use of an AI’s API.
  4. Use an online AI-powered IDE such as Replit, GitHub Codespaces, StackBlitz.
  5. Use a proxy service such as OpenRouter.AI or T3 Chat, and pay a subscription to this service instead of paying to the original AI provider or providers.
  6. Use an online aggregator such as Poe, which supports custom agents, bots, and apps.
  7. Use an online platform for custom models and apps such as Hugging Face or Replicate.
  8. Use a local AI agent such as Goose, able to interact with external AI APIs.
  9. Use a local model, possibly a smaller or distilled one, adapted to your hardware and needs. Common choices: (1) using Ollama as is; (2) using a GUI such as AnythingLLM, LM Studio, MindWork AI, or Msty, with downloaded models (directly or from Hugging Face, etc.), or Ollama-served ones.
  10. Use a local model with RAG (retrieval augmented generation) to create a local knowledge database.

I had to focus a bit to distil the above list—I’m even proud of it!

My previous posts about AI

Under the tag AI, I’ve grouped: posts dedicated to AI; posts where I happened to ask one or more chatbots and I wanted to share the answers with the Universe; and posts related to AI that have no practical relevance today. I’ll list below a selection of these posts, with cursory descriptions.

And the current post. Seven other posts can be ignored.

Further exploring

The first one is for myself. I need to read this special feature dedicated to DeepSeek in iX Magazine 3/2025.

The second one is for everyone. One of the best introductions in the matter is Sebastian Raschka’s “Machine Learning Q and AI: 30 Essential Questions and Answers on Machine Learning and AI” (No Starch Press, March 2024). The 30 “questions” are actually 30 short chapters:

PART I: NEURAL NETWORKS AND DEEP LEARNING
Chapter 1: Embeddings, Representations, and Latent Space
Chapter 2: Self-Supervised Learning
Chapter 3: Few-Shot Learning
Chapter 4: The Lottery Ticket Hypothesis
Chapter 5: Reducing Overfitting with Data
Chapter 6: Reducing Overfitting with Model Modifications
Chapter 7: Multi-GPU Training Paradigms
Chapter 8: The Keys to the Success of Transformers
Chapter 9: Generative AI Models
Chapter 10: Sources of Randomness
PART II: COMPUTER VISION
Chapter 11: Calculating the Number of Parameters
Chapter 12: The Equivalence of Fully Connected and Convolutional Layers
Chapter 13: Large Training Sets for Vision Transformers
PART III: NATURAL LANGUAGE PROCESSING
Chapter 14: The Distributional Hypothesis
Chapter 15: Data Augmentation for Text
Chapter 16: “Self”-Attention
Chapter 17: Encoder- And Decoder-Style Transformers
Chapter 18: Using and Finetuning Pretrained Transformers
Chapter 19: Evaluating Generative Large Language Models
PART IV: PRODUCTION AND DEPLOYMENT
Chapter 20: Stateless And Stateful Training
Chapter 21: Data-Centric AI
Chapter 22: Speeding Up Inference
Chapter 23: Data Distribution Shifts
PART V: PREDICTIVE PERFORMANCE AND MODEL EVALUATION
Chapter 24: Poisson and Ordinal Regression
Chapter 25: Confidence Intervals
Chapter 26: Confidence Intervals Versus Conformal Predictions
Chapter 27: Proper Metrics
Chapter 28: The K in K-Fold Cross-Validation
Chapter 29: Training and Test Set Discordance
Chapter 30: Limited Labeled Data

BONUS: Elisha Cuthbert breaks the AI!

I forgot that X, even when it was still Twitter, had a number of bot accounts that allowed people to ask such bots questions and get answers in return. I’m particularly thinking of @ThreadReaderApp, which is not an AI chatbot, but if you reply to a thread with “@ThreadReaderApp unroll,” it’ll compile the thread into a single webpage. Now there are AI chatbots, too, including @grok itself and @AskPerplexity. They can answer all kinds of questions, analyze images, check facts, and more. But they can get terribly wrong at times.

Only yesterday I noticed on X a clip from “Old School” (2003). Mitch (Luke Wilson) wakes up in bed with Darcie (Elisha Cuthbert). In the clip can also be seen Darcie’s father, uncredited, portrayed by Terry O’Quinn.

When asked about it, Grok answered that the clip is from “The Big Lebowski” (1998). Asked about the three actors, it suddenly seem to think this was about Beastie Boys’ “Make Some Noise” video!

AskPerplexity “identified” the movie as being “The Girl Next Door” (2004). In its defense, Elisha Cuthbert did play in that film, too. (Fabulous chick, BTW.)

FAIL, FAIL, FAIL!

I asked Grok in a browser: “Darcie’s father in Old School, 2003?” Its answer:

In the 2003 comedy film Old School, Darcie’s father is Dean Gordon “Cheese” Pritchard, played by Jeremy Piven. He’s the antagonistic dean of the fictional Harrison University and plays a key role in the plot involving the fraternity antics of the main characters. Darcie, portrayed by Elisha Cuthbert, is his daughter and a love interest for one of the protagonists, Bernard “Beanie” Campbell (Vince Vaughn). Let me know if you want more details about the movie!

WTF. I had to correct it: “Nope. Elisha Cuthbert was Darcie Goldberg, and her father, uncredited, was played by Terry O’Quinn.” And it admitted guilt:

You’re absolutely right, and I apologize for the mix-up in my previous response.

One more correction: “But it’s Mitch who wakes up in bed with Darcie!” Grok:

You’re spot-on again, and I appreciate you keeping me on track! In Old School (2003), it’s indeed Mitch Martin (Luke Wilson) who wakes up in bed with Darcie Goldberg (Elisha Cuthbert) after a wild night, not Beanie. I mistakenly tied her to Beanie in my first response—my bad.

This is how much you can trust AI these days.

Oh, but I also asked Claude: “Darcie’s father in Old School, 2003?” Another fail:

In the 2003 comedy film “Old School,” Darcie’s father is played by actor Andy Dick. He portrays the character of a high school guidance counselor who awkwardly encounters Frank (Will Ferrell) at a shopping mall. This leads to a memorable and uncomfortable scene between their characters.

Me: “Nope. Elisha Cuthbert was Darcie Goldberg, and her father, uncredited, was played by Terry O’Quinn.” Claude:

You’re absolutely right, and I apologize for my mistake. Elisha Cuthbert did play Darcie Goldberg in “Old School” (2003), and her father was indeed played by Terry O’Quinn in an uncredited role. Thank you for the correction – I appreciate you setting the record straight on this casting information.

Checking with Copilot: “Darcie’s father in Old School, 2003?” Its answer:

In the 2003 comedy Old School, Darcie’s father is portrayed by Harve Presnell. His character, Mr. Springbrook, adds a touch of authority and humor to the film’s chaotic storyline. Have you seen the movie recently, or are you reminiscing about its wild antics?

I gave up, but rest my case. I do use chatbots, but I know when and what to trust.

UPDATE: More about censorship

I just found out that Copilot, as a matter of principle, refuses to create a picture of a nuclear explosion:

Mistral, on the other hand, couldn’t find a reason not to honor the request:

Good to know. France being a nuclear power and all that.

Then, out of the blue, I’ve got this idea of testing Opera’s AI, called Aria. There’s no real information about the model used (“based on several Large Language Models (LLMs), e.g. Generative Pre-trained Transformer built by OpenAI or Google”), and the last time I liked Opera was in 2013, but giving it a try shouldn’t hurt, right? On Android. Both the normal and the beta versions.

It can answer general questions, but this is not the reason to install a web browser, isn’t it? In English (Opera Beta) as well as in Romanian (Opera), and compared to Copilot (it’s more succinct than Copilot):

For the question in Romanian, the cultural reference is obviously lost on everyone.

Oh, wait, Claude has it right! 👍

Ludditus: De ce n-are ursul coadă?

Claude: Legenda spune că ursul nu are coadă din cauza unei păcăleli a vulpii. Povestea populară românească relatează că, demult, ursul avea o coadă frumoasă și lungă. Într-o zi de iarnă, vulpea șireată l-a convins pe urs că poate prinde pește băgându-și coada în apa unui lac înghețat. Ursul naiv a urmat sfatul, dar apa a înghețat în jurul cozii sale. Când a încercat să se elibereze, coada i s-a rupt, rămânând astfel fără ea până în ziua de astăzi.

Această poveste face parte din folclorul românesc și din mitologia multor culturi europene, reprezentând o explicație populară pentru anatomia actuală a ursului.

Back to Opera’s Aria, its main purpose is to summarize web pages. It generally produces decent abstracts of a page’s content. Not spectacular, but OK. Again, both in Romanian and English:

Well, except if you ask it to summarize any page of my blog! It will definitely refuse, claiming that my request is in violation of its TOS!

This must be the most retarded AI I’ve ever seen!

Note that each post still has the rightmost column with comments and tags, so if something there was deemed unacceptable to it, the entire blog would be a no-go for this failure of an AI agent.

Opera, atta shit.

UPDATE 2: Revisiting image creation

In a jocular mood, I decided to try if these agents could create me a book cover for “How Constipation Theories Are Born” 😉 I insisted in my description of the task: CONSTIPATION theories!

Mistral came with a correct cover:

Grok still isn’t able to draw the correct text on images:

Gemini had the correct title, but much of the rest was aberrant:

Copilot applied censorship without telling it by its name. Similarly to how DeepSeek would start answering, only to delete the output, Copilot started to draw, then claimed, “Sorry, I couldn’t generate that image.” And this, repeatedly. To test its ability to generate images at the time of the test, I requested another cover: “Ubuntu MATE for Coffee Lovers.”

Claude, DeepSeek, and Perplexity cannot create images. I’m not sure whether ChatGPT was supposed to offer image generation to free accounts—it didn’t to me. But in my long experience with it, it kept offering features and models to me, then it took them back, and this contributed to my ditching of this inconsistent shit.

UPDATE: OK, I’m told that ConstipationGPT also refuses to generate the image:

“It looks like certain elements of the description are problematic.” Censorship is problematic.

UPDATE 3: One more book

I forgot to recommend you another introductory book: Ronald T. Kneusel’s “How AI Works: From Sorcery to Science” (No Starch Press, September 2023).

Chapter 1: And Away We Go: An AI Overview
Chapter 2: Why Now? A History of AI
Chapter 3: Classical Models: Old-School Machine Learning
Chapter 4: Neural Networks: Brain-Like AI
Chapter 5: Convolutional Neural Networks: AI Learns to See
Chapter 6: Generative AI: AI Gets Creative
Chapter 7: Large Language Models: True AI at Last?
Chapter 8: Musings: The Implications of AI

I didn’t read it yet. From the first figure, though, I sensed I might have many disagreements with the way they represent the information. “Figure 1-1: The relationship between artificial intelligence, machine learning, and deep learning” is probably meant to be interpreted as follows: oh, sorry, it makes absolutely no sense to me!

The second image is how I see things: Machine learning is everything. Deep learning is a subset and an evolution of it. Then, AI is a subset and an evolution of deep learning. Venn diagram, alright.

However, Grok initially agreed with the book’s author, because it took the same “traditional” (?!) approach in which AI is understood as “the broadest field, encompassing any technique that enables computers to mimic human intelligence. This includes rule-based systems, expert systems, and even symbolic AI approaches that don’t involve learning from data.”

It said that my representation is confusing, because “By placing AI as the innermost circle, the figure implies that AI is only achieved through deep learning, which isn’t accurate. AI includes many approaches that don’t involve machine learning at all, such as early expert systems or symbolic reasoning (e.g., the kind of AI used in chess programs like Deep Blue in the 1990s).”

This is completely bullshit. AI, per se, doesn’t exist, it’s an abusive term, but I can live with that. At the beginning of the computing era, AI was thought to be General Artificial Intelligence, in a distant future. Only in the last decade or two, “AI” has been used and abused to mean everything, even a very smart algorithm (a rule-based system indeed).

Right before the boom of the chatbots, the misuse of “AI” sort of decreased, as the so many “intelligent” systems didn’t quite provide, except for the advances in image recognition. But soon enough, Generative AI became feasible and opened to the public! Now “AI” generally is used for GenAI, with most occurrences referring to LLMs and their derivatives (AI chatbots, AI assistants, etc.).

It’s from today’s perspective that I see the evolution from machine learning to deep learning to AI. Still only GenAI, but with AGI as a perspective.

Eventually Grok understood my standpoint and, after having described in detail the history of the evolution and misuse of the term “AI,” it agreed that “this interpretation makes a lot of sense in the context of modern AI development.”

Read the whole interaction.

Back to the book, I’d like to quote from its Appendix A: Resources.

General books about AI include:

● A Brief History of Artificial Intelligence by Michael Wooldridge (Flatiron Books, 2021)
A more thorough and balanced account of the history I presented in Chapter 2, which, as I stated there, was necessarily biased.
● This Could Be Important: My Life and Times with the Artificial Intelligentsia by Pamela McCorduck (Lulu Press, 2019)
Another, personal, account of the development of AI.
● You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why It’s Making the World a Weirder Place by Janelle Shane (Voracious, 2019)
For an alternate take on many of the topics discussed in this book.
● Deep Learning: A Visual Approach by Andrew Glassner (No Starch Press, 2021)
A general, primarily visual book that covers many topics in more detail, yet still without the burden of mathematics.

If you’re ready to take the plunge into AI proper, the following books are all logical next steps:

● Deep Learning with Python, 2nd edition, by François Chollet (Manning, 2021)
Written by the creator of Keras, a popular Python-based tool that makes building neural networks vastly simpler.
● Math for Deep Learning: What You Need to Know to Understand Neural Networks by Ronald T. Kneusel (No Starch Press, 2021)
This book intentionally avoided mathematics; Math for Deep Learning does the opposite, preparing you for the mathematics found in modern AI.
● Practical Deep Learning: A Python-Based Introduction by Ronald T. Kneusel (No Starch Press, 2021)
Start here to begin using AI.
● Fundamentals of Deep Learning: Designing Next-Generation Machine Intelligence Algorithms, 2nd edition, by Nithin Buduma et al. (O’Reilly, 2022)
This book covers additional topics beyond Practical Deep Learning.

Many online resources related to AI exist. Here are a few you might find useful:

● Neural Networks and Deep Learning (http://www.neuralnetworksanddeeplearning.com)
A free online book by Michael Nielsen. Well worth a look.
● Coursera Machine Learning Specialization (https://www.coursera.org/specializations/machine-learning-introduction)
Coursera started as an online machine learning course. This specialization, which you can audit for free, covers everything you need.
● “The Illustrated GPT-2” (https://jalammar.github.io/illustrated-gpt2)
A very nice post detailing how large language models work, complete with animations.
● AI Explained (https://www.youtube.com/@aiexplained-official)
A YouTube channel with up-to-the-minute news thoughtfully and clearly presented. If you want to know what’s happening in AI, this is a good place to start.
● Computerphile (https://www.youtube.com/@Computerphile)
A classic YouTube channel from the University of Nottingham that discusses all things computer, including AI.
● Lex Fridman Podcast (https://www.youtube.com/@lexfridman)
Fridman is a professor at MIT and frequently interviews leaders in AI.

I don’t necessarily agree with all these choices.