Every few months I put together a guide on which AI system to use. Since I last wrote my guide, however, there has been a subtle but important shift in how the major AI products work. Increasingly, it isn't about the best model, it is about the best overall system for most people. The good news is that picking an AI is easier than ever and you have three excellent choices. The challenge is that these systems are getting really complex to understand. I am going to try and help a bit with both.

First, the easy stuff.

Which AI to Use

For most people who want to use AI seriously, you should pick one of three systems: Claude from Anthropic, Google’s Gemini, and OpenAI’s ChatGPT. With all of the options, you get access to both advanced and fast models, a voice mode, the ability to see images and documents, the ability to execute code, good mobile apps, the ability to create images and video (Claude lacks here, however), and the ability to do Deep Research. Some of these features are free, but you are generally going to need to pay $20/month to get access to the full set of features you need. I will try to give you some reasons to pick one model or another as we go along, but you can’t go wrong with any of them.

What about everyone else? I am not going to cover specialized AI tools (some people love Perplexity for search, Manus is a great agent, etc.) but there are a few other options for general purpose AI systems: Grok by Elon Musk’s xAI is good if you are a big X user, though the company has not been very transparent about how its AI operates. Microsoft’s Copilot offers many of the features of ChatGPT and is accessible to users through Windows, but it can be hard to control what models you are using and when. DeepSeek r1, a Chinese model, is very capable and free to use, but is missing a few features from the other companies and it is not clear that they will keep up in the long term. So, for most people, just stick with Gemini, Claude, or ChatGPT

Great! This was the shortest recommendation post yet! Except… picking a system is just the beginning. The real challenge is understanding how to use these increasingly complex tools effectively.

Now what?

I spend a lot of time with people trying to use AI to get stuff done, and that has taught me how incredibly confusing this is. So I wanted to walk everyone through the most important features and choices, as well as some advice on how to actually use AI.

Picking a Model

ChatGPT, Claude, and Gemini each offer multiple AI models through their interface, and picking the right one is crucial. Think of it like choosing between a sports car and a pickup truck; both are vehicles, but you'd use them for very different tasks. Each system offers three tiers: a fast model for casual chat (Claude Sonnet, GPT-4o, Gemini Flash), a powerful model for serious work (Claude Opus, o3, Gemini Pro), and sometimes an ultra-powerful model for the hardest problems (o3-pro, which can take 20+ minutes to think). The casual models are fine for brainstorming or quick questions. But for anything high stakes (analysis, writing, research, coding) usually switch to the powerful model.

Most systems default to the fast model to save computing power, so you need to manually switch using the model selector dropdown. (The free versions of these systems do not give you access to the most powerful model, so if you do not see the options I describe, it is because you are using the free version)

I use o3, Claude 4 Opus, and Gemini 2.5 Pro for any serious work that I do. I also have particular favorites based on individual tasks that are outside of these models (GPT-4.5 is a really interesting model for writing, for example), but for most people, stick with the models I suggested most of the time.

For people concerned about privacy, Claude does not train future AI models on your data, but Gemini and ChatGPT might, if you are not using a corporate or educational version of the system. If you want to make sure your data is never used to train an AI model, you can turn off training features easily for ChatGPT without losing any functionality, and at the cost of some functionality for Gemini. You may also want to turn on or off “memory” in ChatGPT’s personalization option, which lets the AI remember scattered details about you. I find the memory system to be too erratic at this point, but you may have a different experience.

Using Deep Research

Deep Research is a key AI feature for most people, even if they don’t know it yet. Deep Research tools are very useful because they can produce very high-quality reports that often impress information professionals (lawyers, accountants, consultants, market researchers) that I speak to. You should be trying out Deep Research reports in your area of expertise to see what they can do for you, but some other use cases include:

• Gift Guides: “what do I buy for a picky 11-year-old who has read all of Harry Potter, is interested in science museums, and loves chess? Give me options, including where to buy at the best prices.”

• Travel Guides “I am going to Wisconsin on vacation and want to visit unique sites, especially focusing on cheese, produce a guide for me”

• Second opinions in law, medicine, and other fields (it should go without saying that you should trust your doctor/lawyer above AI, but research keeps finding that the more advanced AI systems do very well in diagnosis with a surprisingly low hallucination rate, so they can be useful for second opinions).

Deep Research reports are not error-free but are far more accurate than just asking the AI for something, and the citations tend to actually be correct. Also note that each of the Deep Research tools work a little differently, with different strengths and weaknesses. Turning on the web search option in Claude and o3 will get them to work as mini Deep Research tools, doing some web research, but not as elaborately as a full report. Google has some fun additional options once you have created a report, letting you turn it into an infographic, a quiz or a podcast.

An Easy Approach to AI: Voice Mode

An easy way to use AI is just to start with voice mode. The two best implementations of voice mode are in the Gemini app and ChatGPT’s app and website. Claude’s voice mode is weaker than the other two systems. What makes voice mode great is that you can just have a natural conversation with the app while in the car or on a walk and get quite far in understanding what these models can do. Note the models are optimized for chat (including all of the small pauses and intakes of breath designed to make it feel like you are talking to a person), so you don’t get access to the more powerful models this way. They also don’t search the web as often which makes them more likely to hallucinate if you are asking factual questions: if you are using ChatGPT, unless you hear the clicking sound at 44 seconds into this clip, it isn’t actually searching the web.

Voice mode's killer feature isn't the natural conversation, though, it's the ability to share your screen or camera. Point your phone at a broken appliance, a math problem, a recipe you're following, or a sign in a foreign language. The AI sees what you see and responds in real-time. I've used it to identify plants on hikes, solve a problem on my screen, and get cooking tips while my hands were covered in flour. This multimodal capability is genuinely futuristic, yet most people just use voice mode like Siri. You're missing the best part.

Making Things for You: Images, Video, Code, and Documents

ChatGPT and Gemini will make images for you if you ask (Claude cannot). ChatGPT offers the most controllable image creation tool, Gemini uses two different image generation tools, Imagen, a very good traditional image generation system, and a multimodal image generation system. Generally, ChatGPT is stronger. On video creation, however, Gemini’s Veo 3 is very impressive, and you get several free uses a day (but you need to hit the Video button in the interface)

All three systems can produce a wide variety of other outputs, ranging from documents to statistical analyses to interactive tools to simulations to simple games. To get Gemini or ChatGPT to do this reliably, you need to select the Canvas option when you want these systems to run code or produce separate outputs. Claude is good at creating these sorts of outputs on its own. Just ask, you may be surprised what the AI systems can make.

Working with an AI

Now that you have picked a model, you can start chatting with it. It used to be that the details of your prompts mattered a lot, but the most recent AI models I suggested can often figure out what you want without the need for complex prompts. As a result, many of the tips and tricks you see online for prompting are no longer as important for most people. At the Generative AI Lab at Wharton, we have been trying to examine prompting techniques in a scientific manner, and our research has shown, for example, that being polite to AI doesn’t seem to make a big difference in output quality overall1. So just approach the AI conversationally rather than getting too worried about saying exactly the right thing.

That doesn’t mean that there is no art to prompting. If you are building a prompt for other people to use, it can take real skill to build something that works repeatedly. But for most people you can get started by keeping just a few things in mind:

• Give the AI context to work with. Most AI models only know basic user information and the information in the current chat, they do not remember or learn about you beyond that. So you need to provide the AI with context: documents, images, PowerPoints, or even just an introductory paragraph about yourself can help - use the file option to upload files and images whenever you need. The AIs can do some of these ChatGPT and Claude can access your files and mailbox if you let them, and Gemini can access your Gmail, so you can ask them to look up relevant context automatically as well, though I prefer to give the context manually.

• Be really clear about what you want. Don’t say “Write me a marketing email,” instead go with “I'm launching a B2B SaaS product for small law firms. Write a cold outreach email that addresses their specific pain points around document management. Here's the details of the product: [paste]” Or ask the AI to ask you questions to help you clarify what you want.

• Give it step-by-step directions. Our research found this approach, called Chain-of-Thought prompting, no longer improves answer quality as much as it used to. But even if it doesn’t help that much, it can make it easier to figure out why the AI came up with a particular answer.

• Ask for a lot of things. The AI doesn’t get tired or resentful. Ask for 50 ideas instead of 10, or thirty options to improve a sentence. Then push the AI to expand on the things you like.

• Use branching to explore alternatives. Claude, ChatGPT, and Gemini all let you edit prompts after you have gotten an answer. This creates a new “branch” of the conversation. You can move between branches by using the arrows that appear after you have edited an answer. It is a good way to learn how your prompts impact the conversation.

Troubleshooting

I also have seen some fairly common areas where people get into trouble:

• Hallucinations: In some ways, hallucinations are far less of a concern than they used to be, as AI has improved and newer AI models are better at not hallucinating. However, no matter how good the AI is, it will still make errors and mistakes and still give you confident answers where it is wrong. They also can hallucinate about their own capabilities and actions. Answers are more likely to be right when they come from the bigger, slower models, and if the AI did web searches. The risk of hallucination is why I always recommend using AI for topics you understand until you have a sense for their capabilities and issues.

• Not Magic: You should remember that the best AIs can perform at the level of a very smart person on some tasks, but current models cannot provide miraculous insights beyond human understanding. If the AI seems like it did something truly impossible, it is probably not actually doing that thing but pretending it did. Similarly, AI can seem incredibly insightful when asked about personal issues, but you should always take these insights with a grain of salt.

• Two Way Conversation: You want to engage the AI in a back-and-forth interaction. Don’t just ask for a response, push the AI and question it.

• Checking for Errors: The AI doesn’t know “why” it did something, so asking it to explain its logic will not get you anywhere. However, if you find issues, the thinking trace of AI models can be helpful. If you click “show thinking” you can find out what the model was doing before giving you an answer. This is not always 100% accurate (you are actually getting a summary of the thinking) but is a good place to start.

Your Next Hour

So now you know where to start. First, pick a system and resign yourself to paying the $20 (the free versions are demos, not tools). Then immediately test three things on real work: First, switch to the powerful model and give it a complex challenge from your actual job with full context and have an interactive back and forth discussion. Ask it for a specific output like a document or program or diagram and ask for changes until you get a result you are happy with. Second, try Deep Research on a question where you need comprehensive information, maybe competitive analysis, gift ideas for someone specific, or a technical deep dive. Third, experiment with voice mode while doing something else — cooking, walking, commuting — and see how it changes your ability to think through problems.

Most people use AI like Google at first: quick questions, no context, default settings. You now know better. Give it documents to analyze, ask for exhaustive options, use branching to explore alternatives, experiment with different outcomes. The difference between casual users and power users isn't prompting skill (that comes with experience); it's knowing these features exist and using them on real work.

Subscribe now

Share

There is an inherent problem with the design of shower heads. Not some of them, all of them. The problem is that their very design creates the ideal circumstances for mold to thrive within them, internally, in areas that you cannot access for cleaning.

A bathtub faucet or kitchen sink tap is simply just a shaped pipe that allows water to flow through them. When you turn the water off, the pipe mouths quickly dry, thanks to their relatively wide shape and local airflow.

Showerheads, however, are complex workings of intricate inner channels and nozzles, designed to break the water flow into spray patterns that end users find desirable.

These channels are all inside the showerhead and get little airflow.

The channels can never really dry out completely, and over time, that interal dampness allows bacteria and mold—including the dreaded black mold--to thrive. In this shot of a showerhead that has been cut open by a saw, a lot of what you see is the detritus of the cut plastic, but you can also see the brown stuff.

And deeper inside the head, you find this:

The mother of Steve Sunshine, an inventor, was suffering from respiratory issues. Sunshine disassembled her showerhead and found it was filled with mold. He subsequently designed this ShowerClear:

This ingenious design pops open, so that after a shower you can let the shower head's innards dry out. It also makes it easy to clean, so you can eliminate mineral build-up. (This eliminates the mild hassle that many of us undertake to clean our showerheads, which is soaking them in a vessel filled with vinegar for a few hours.)

The ShowerClear heads come in a variety of finishes and run $140.

Amazon is making it easier for sellers to quickly create generative AI ads on its platform, sometimes with just a single click. Amazon Ad’s Video Generator, a free advertising tool introduced in beta last year, now has some new tricks and is generally available for sellers in the US to create “photorealistic video assets” in five minutes or less.

New capabilities include motion improvements to show items in action, which Amazon says is best for showcasing products like toys, tools, and worn accessories. For example, Video Generator can now create clips that show someone wearing a watch on their wrist and checking the time, instead of simply displaying the watch on a table. The tool generates six different videos to choose from, and allows brands to add their logos to the finished results.

The Video Generator can now also make ads with multiple connected scenes that include humans, pets, text overlays, and background music. The editing timeline shown in Amazon’s announcement video suggests the ads max out at 21 seconds.. The resulting ads edge closer to the traditional commercials we’re used to seeing while watching TV or online content, compared to raw clips generated by video AI tools like OpenAI’s Sora or Adobe Firefly.

A new video summarization feature can create condensed video ads from existing footage, such as demos, tutorials, and social media content. Amazon says Video Generator will automatically identify and extract key clips to generate new videos formatted for ad campaigns. A one-click image-to-video feature is also available that creates shorter GIF-style clips to show products in action.

We’ll likely see Amazon retailers utilizing AI-generated video ads in the wild now that the tool is generally available in the US and costs nothing to use — unless the ads are so convincing that we don’t notice anything at all.

The difference can be subtle

As you progress in your career, understanding the difference between being useful and being valued is very important. At first glance, they might look similar because the signals you get are more or less the same: a promotion, a higher than expected bonus, a special stock award. This is why it’s important to dig deeper and try to detect subtler signals.

Being useful means that you are good at getting things done in a specific area, so that people above you can delegate that completely. You are reliable, efficient, maybe even indispensable in the short term. But you are seen primarily as a gap-filler, someone who delivers on tasks that have to be done but are not necessarily a core component of the company strategy. “Take care of that and don’t screw up” is your mission, and the fewer headaches you create for your leadership chain, the bigger the rewards.

Being valued, on the other hand, means that you are brought into more conversations, not just to execute, but to help shape the direction. This comes with opportunities to grow and contribute in ways that are meaningful to you and the business.

It took me a few years to truly grasp the difference. If you’re valued, you’ll likely see a clear path for advancement and development, you might get more strategic roles and involvement in key decisions. If you are just useful, your role might feel more stagnant.

Let me walk you through two contrasting experiences from my own career that made this difference clearer to me.

Feeling valued during layoffs

When I was an IC, there was a period when my company was going through a tough time. We missed earnings guidance by a lot in one quarter, and layoffs were inevitable. Entire teams were disbanded, and my own manager was among those let go, so I was quite nervous. But instead of being on the chopping block, I was not only retained but also offered a significant retention bonus: 50% of my total comp, with a one year vesting schedule.

Leadership made it clear that they saw me as critical to the company’s future, not just because of what I had delivered in the past, but to help shape what came next. The business was changing, becoming more digital, and they needed the skills I was bringing to the table. Interestingly, this validation didn’t come from a performance review or a bonus alone. It came in a moment of crisis, when hard choices had to be made about who mattered most to the mission ahead.

The rewards of being useful

Later on, I found myself in a role that, on the surface, looked like a success story. I was consistently hitting targets, getting recognition from leadership, and receiving generous bonuses and retention grants. The company clearly saw me as someone worth keeping around. But over time, I started to notice that while I was being rewarded, I wasn’t really being asked to take on new kinds of problems or invited into strategic conversations. No one was talking to me about where I wanted to go next or how I could grow.

I had become the go-to person for making things run smoothly, for fixing urgent problems, for delivering. But every time I pushed toward more strategic and ambitious directions, there was a lot of can-kicking and “let’s think about it” that went nowhere. I was incredibly useful to the organization, but not necessarily valued, and at some point, I started feeling a sense of stagnation. Compensation was good, the actual job was aligned with my interests, but that sense of being just a useful caretaker was hitting my motivation. In the end, I had to move on to another role.

If you’re reading this and wondering which side of the line you are on, I encourage you to take a moment to step back and look beyond the surface. Are you valued, or just useful?

---

Adblock test (Why?)

I’ve been involved in content moderation since roughly 2004. I’ve built spam prevention for corporate and personal e-mail, moderated open-source mailing lists and IRC channels, worked at a couple social media networks, and help moderate a Mastodon instance for a few hundred people. In the last few years I’ve wasted more time fighting blog comment spam, and I’m pretty sure Large Language Models (LLMs) are to blame.

I think of spam as a space with multiple equilibria. Producing spam takes work. Spammers are willing to invest that work because each message has a small chance to make money, or achieve political or emotional goals. Some spam, like the endless identical Viagra scams in my email spam folder, or the PHPBB comment spam I filter out here on aphyr.com, is cheap to generate and easy to catch. I assume the spammers make it up in volume. Other spam, like spear phishing attacks, is highly time-consuming: the spammer must identify a target, carefully craft a plausible message using, say, the identity of the target’s co-workers, or construct a facade of a bank’s log-in page, and so on. This kind of spam is more likely to make it through filters, but because it takes a lot of human work, is generally only worth it for high-value targets.

LLMs seem to be changing these equilibria. Over the last year I’ve seen a new class of comment spam, using what I’m fairly sure is LLM-generated text. These comments make specific, plausible remarks about the contents of posts and images, and work in a link to some web site or mention a product. Take this one I caught a few months back:

"Walking down a sidewalk lined with vibrant flowers is one of life’s simple joys! It reminds me of playing the [link redacted] slope game, where you have to navigate through colorful landscapes while dodging obstacles.

Before 2023, you’d likely have paid a human a few cents to write and post that. Now, thanks to LLMs, this sort of thing is trivially automated. The model will happily fabricate relatable personal experiences in service of a spam campaign:

That photo reminds me of the first time I tried macro photography in my backyard. I spent an hour trying to get a clear shot of a red flower, experimenting with angles and lighting. It was so much fun discovering the little details up close! If you ever need a break from photography, I recommend playing Snow Rider 3D for a bit of quick, light-hearted fun.

Other spam seems to glue together LLM remarks with what I think is a hand-written snippet of ad copy. Note the abrupt shift in grammar, diction, and specificity.

This piece masterfully blends technical depth with mythological storytelling, transforming a standard Haskell programming interview into an epic narrative. It cleverly critiques the complexity and absurdity of some technical interviews by illustrating how type-level Haskell can be pushed to esoteric extremes beautiful, powerful, and largely impractical. A fascinating and relevant read for anyone interested in the intersection of programming, language design, and narrative. I’m James Maicle, working at Cryptoairhub where we run a clear and insightful crypto blog. I’ll be bookmarking your site and following the updates. Glad to see so much valuable information shared here looking forward to exploring more strategies together. Thanks for sharing. If you interest about Crypto please visit my website and read my article [link redacted] Crypto Blog.

The same thing is happening on aggregators like Hacker News, where commenters post more-or-less-obviously LLM output for… I’m not sure, exactly. Karma? Weirder still are bots like Hacker Briefs, which I suspect use an LLM to summarize trending HN posts. Here’s its summary of a recent article I wrote:

“Jepsen: Amazon RDS for PostgreSQL 17.4”

New multi-AZ clusters in Amazon RDS for PostgreSQL offer better failure recovery but may return outdated data when reading after writes. Caution is needed.

This is a totally plausible summary of the article, and it is utterly, laughably wrong. Multi-AZ clusters are not new, and they do not return outdated data when reading after writes. As the abstract succinctly explains, they allow Long Fork, a different anomaly which does not involve real-time orders at all. The bot ignored the actual problem and invented a different one. This sort of spam isn’t obviously motivated by commercial interest, but it is nevertheless depressing: one more drop in the misinformation slurry.

Of course this is not news. Product reviews are inundated with LLM slop, as are social media networks. LLMs allow for cheap, fast, and automated generation of unique spam which is difficult for machines and humans to identify. The cost falls on me and other moderators, who must sift through LLM bullshit trying to sieve “awkward but sincere human” from “automated attack”. Thanks to OpenAI et al I read more spam, and each message takes longer to check.

This problem is only going to get worse as LLMs improve and spammers develop more sophisticated ways to use them. In recent weeks I’ve received vague voice messages from strangers with uncanny speech patterns just asking to catch up—a sentence which, had I uttered it prior to 2023, would have been reasonably interpreted as a sign of psychosis. I assume these too are LLM-generated scams, perhaps a pig butchering scheme. So far these are from strangers, but it’s not hard to imagine an attacker using text and voice synthesis to impersonate a friend, colleague, or lover in a detailed conversation. Or one’s doctor, or bank.

As the cost of generating slop decreases, it’s easy to imagine LLMs generating personae, correspondence, even months-long relationships with real humans before being deployed for commercial or political purposes. Creating plausible accounts and selling them has been a successful business model in social media for some time; likewise, we have strong clues that LLMs are already used for social media bots. Social networks have responded to these attacks via out-of-band mechanisms: IP reputation analysis, javascript and mobile app fingerprinting, statistical correlation across multiple accounts, and so on. I’m not sure how to translate these defenses to less centralized and more privacy-oriented networks, like email or blog spam. I strongly suspect the only reason Mastodon hasn’t been eaten alive by LLM spambots is because we’re just not big enough to be lucrative. But those economics are shifting, and even obscure ecological niches can be worth filling.

As a moderator, that keeps me up at night.

Adblock test (Why?)

Source - Patreon