Important (2025.4.18)

Big-company Google is awesome. US Edu email verification gives free 15 months directly. Straight up freebies! via: https://gemini.google/students/

Enough said—Google rocks! Looking forward to Gemini eventually sweeping all LLMs, dominating alone!

Update [2025/4/12]

At this point [April 12, 2025], Gemini Advanced is absolutely worth paying for. See my blog post Gemini 2.5 Pro Experience. If too expensive, can buy monthly throwaway accounts from Taobao or Xianyu—saw some L Site folks getting Gemini Advanced this way cheaply.

Update [2025/1/30]

Combined with past months’ experience, Gemini Advanced’s main selling point is using Gemini 1.5 Pro to summarize YouTube videos. Everything else feels lacking—experience isn’t as good as Google AI Studio. Plus video summarization is based on subtitles, not strong multimodal capabilities understanding video. Below I mentioned summarizing a no-subtitle video—that was inaccurate. No-subtitle videos and Trump-related videos basically can’t be summarized.

Update [2025/1/9]

Gemini’s latest Exp models recommended to try free in AI Studio first—Gemini Advanced’s experimental models are still lacking.

I renewed Gemini Advanced mainly for Google One benefits, plus occasionally using Gemini Advanced’s 1.5 Pro to quickly summarize YouTube videos.

Deep Research isn’t frequently used either—after all, lots of valuable knowledge is trapped in “walled garden” apps. Public internet has lots of garbage content, especially after generative AI explosion in 2022.

Prioritize gemini-exp-1206 model—after all, it’s Gemini 2.0 Pro preview.

Image recognition is impressive. Upload large images—text inside can be recognized nearly accurately. Gemini’s future is promising.

Looking forward to Gemini 2.0 full release around Lunar New Year!

Update [2024/12/15]

Google One AI Premium is increasingly worthwhile, though Google AI Studio is free. Next year NotebookLM membership will also join Google One AI Premium package—this $20 is worth it.

Making a prediction: Google Gemini 2.0 full release early next year, its lead in AI track will become apparent. Within next 10 years, the closest to AGI form will be Gemini. “Started early, arrived late” criticism of Google as a loser may get slapped in the future.

December’s Google Gemini 2.0 release was amazing—much better than OpenAI’s 12 days of 10-20 minute daily video demos. Too busy lately to write review notes—anyway, it’s very strong.

For coding questions, prioritize Claude and Google AI Studio’s gemini-exp-1206.

Update [2024/11/16]

Renewed because my shared Turkish Google One+YouTube Premium fell through recently [Google’s cross-region crackdown hammer finally hit me]. I think this $20 is mainly for Google One benefits, Gemini App as bonus.

Though Gemini is destined to not give me Claude-like coding experience.

---

Note: This post published November 2, 2024. After reflection decided I can’t delay anymore—otherwise next month when Gemini has another big update, this article becomes worthless.

Update [2024/11/1, 2024/11/2]

• 2024/11/1

Last night Gemini product updated again. Below experience is from second half of October until before the update.

• 2024/11/2

Yesterday (2024/11/1) tried Grounding feature in Google Studio, also called it in Colab. My feeling: current Grounding works better with English. Google searching Chinese returns too much low-quality content.

English replies all correct, Chinese reply had wrong Tatum score.

Can’t search my blog yet.

Probably my site’s weight isn’t enough—ranking too low on Google caused Grounding response to return empty array.

Personally I think RAG applications adding Google Grounding search path [adding external search] is a good choice—after all, Google is the search engine leader.

Of course Google Grounding will still collect user data like Gemini API free tier. Don’t worry too much—data is LLM’s nourishment, every tech giant sacrifices user privacy for stronger LLM. Like Elon Musk’s X, starting November 15, will use all user data to train Grok, without opt-out option like before.

Initially I wanted to delete X account, but later thought—almost all tech giants use user data to train LLM, some openly, some secretly. Training is training—I’ve somewhat given up on privacy. In LLM era, privacy can hardly stay intact. Just keep sensitive info from being used—public personal data doesn’t matter.

Below formatting might be messy [timeline confused]. During my half-month Gemini Advanced experience, I recorded thoughts in Yuque whenever I had them—formatting inevitably got messy.

Also, things mentioned below like code comments, Kun Kun puzzle aren’t representative—can’t well reflect model capability. Just my momentary thoughts during experience.

Sometimes model’s stupid replies might also be because I was lazy with prompts, failing to get ideal replies. When using LLM, if you want satisfactory replies, questions must clearly describe your needs.

---

API Drawbacks

Paid API isn’t used for model training.

Free API will be used to improve products.

via: https://ai.google.dev/gemini-api/terms

See Simonwillison’s tweet on Gemini privacy issues.

Don’t think Google is benevolent—API is free but they use your conversations for training. As this comment says: if you’re not paying for the product, you are the product. Hits current internet reality—whether using domestic or foreign products like Baidu or Google, while enjoying free quota, your data (aka nourishment) is used to improve products.

Could it be that dumb Gemini Advanced in the morning, suddenly smart in the afternoon because of manual review? Thinking about it is quite unsettling.

Trust it once—turning off activity won’t feed my data to AI training. Anyway this extension is optional for me.

Compared to other products like ChatGPT or Claude, they can save chat history and opt out of training—unclear if they’ll actually follow user agreements. In AI wave, whoever controls data controls first-mover advantage. Google just doesn’t pretend—either give your chat history for improving AI, or don’t save chat history.

Actually strange—some foreigners dislike Google, think it’s evil with privacy issues, while some Chinese folks seem fond of Google, going All in Google—Pixel garbage phones, Google One AI Premium, YouTube Premium all subscribed. Think products are good, but after using find Google also has privacy issues, using your data to improve their products.

Privacy and convenience conflict—indeed true. I’ve turned off activity for now—anyway useful replies I save immediately with notes and screenshots.

Update:

Turned it back on—might need it later. Plus in this AI wave, your data will be used to train AI—like Musk’s X, domestic companies too. ByteDance’s PixelDance surely won’t spare Douyin or TikTok’s massive videos. Sigh 😑, accept the fate of being trained—or if upset, create X alt account to pollute Grok training data.

Look at LinkedIn’s updated user agreement—also starting to use user data to train AI.

Usage Tips

For Chinese internet questions, just ask in Chinese. For technical or questions needing foreign web searches, suggest asking in English.

Or just ask everything in English, then use immersive translate for English replies. After all, using Google’s Double Check in English leads to more reliable sites—you probably don’t want Google’s Double Check going to CSDN 😁.

Highlights

Gemini Advanced referral lets friends get 4 months free.

$9.9 for 5 months total experience—main account subscribes first month, cancels second month, near end of first month invite your alt account, alt gets 4 months Google One AI Premium benefits. Sweet. But if this month’s experience is good, I’ll continue subscribing next month.

Highly integrated with Google ecosystem. Too many YouTube videos to watch, no time—just get the summary. Have Gemini summarize YouTube videos.

Summarizing YouTube video summaries—Gemini API doesn’t support, only Gemini app supports.

In this information overload era, I increasingly can’t calm down to appreciate creators’ long videos. Future non-hands-on YouTube videos—I’ll probably mostly have Gemini Advanced summarize first.

Can summarize YouTube videos without membership too—just not using the most advanced model. Summaries are about getting the gist—detailed or brief summaries are similar.

Even can summarize instructor’s beautiful life—must admire Gemini 1.5 Pro’s video processing ability.

Note: original instructor video has no CC subtitles. Multimodal Gemini is impressive.

In one go had Gemini summarize basically all Watch Later YouTube videos—left 1-2 hands-on videos for when I have time. Comfortable, great!

Indeed Watch Later = Never Watch—where’s the energy? 24 hours a day, knowledge you can’t digest now—you expect to savor it in the future?

Largest context window.

Baffling

After morning Gemini Advanced experience, somewhat disappointed with Google. After afternoon nap, I directly logged out of Google account, re-logged in, then found Gemini Advanced’s intelligence returned. Maybe after subscribing to membership, initially doesn’t provide latest model.

Kun puzzle solvable.

Code comments can also complete in sections.

Weak Points

• Politically sensitive

Whether summarizing YouTube video summaries or regular US election questions—all refused. Conservative Google fears making mistakes on election.

• Supported file types not as good as GPT, Claude

Like not supporting .ipynb file upload.

Uploading Colab .ipynb file from Google Drive—directly shows folder empty.

Actually Colab Notebooks folder isn’t empty.

Of course, criticizing Gemini for not supporting .ipynb is excessive. After all, GPT 4o and Claude are also dumb.

GPT 4o’s comment version—every line comment is the same.

Claude’s is also defective.

Looks like LLM’s .ipynb file parsing still has room for improvement.

Solution is to convert .ipynb to py file first, then upload to LLM for comments.

You think Gemini Advanced can write great comments?

Wrote a small section then truncated.

In contrast, API console—wow, really presented all line comments.

Claude writing comments exceeded context window. But can continue.

If using Google’s Double Check, don’t ask in Chinese, especially programming questions—probably checks CSDN sites.

Gemini dumb replies:

Sometimes I start doubting how Gemini handles context—Chinese understanding is lacking.

For Chinese comments, still need GPT. Claude and Gemini are both lazy.

Gemini Advanced is simply an idiot.

Look at o1-mini’s reply—so complete.

Imagen3 can’t generate people? I remember free version doesn’t support generating people, Advanced version does.

Dumb Gemini can’t even access papers.

Can only upload one image at a time—limits my expression.

Console is way better than Gemini App. Garbage Gemini. Google doesn’t put effort into Gemini App.

Gemini comprehension needs improvement.

Changed question phrasing—responded normally again.

Other

Saw an interesting discussion on Reddit: https://www.reddit.com/r/Bard/comments/1g8ktp7/Gemini_is_the_worst_ai_product_in_the_market/

Outlook

Cancelled subscription. Cancelled both Gemini and ChatGPT [Note: After Search release, re-subscribed to GPT—after all, OpenAI’s second anniversary at month end might have surprises]. Wait until year end—if there’s a trump card, re-subscribe to ChatGPT. Gemini App really unnecessary to subscribe—although includes Google One plan, Gemini App is extremely lacking. Better to use console for free [or subscribe for powerful model to summarize YouTube, enjoy Gemini’s most advanced model while also enjoying Google One benefits. This $20/month is quite worth it—though model may be slightly weaker than Claude, GPT, possibly stronger than 4o].

via: https://www.theverge.com/2024/10/25/24279600/Google-next-Gemini-ai-model-openai-december

As year end approaches, increasingly looking forward to OpenAI, Gemini, even Elon Musk’s Grok bringing more surprises.