People still write?

This is my decision tree for which model to use on #ChatGPT right now.

𝗢𝟯: Use by 𝗱𝗲𝗳𝗮𝘂𝗹𝘁.
𝗢𝟰-𝗺𝗶𝗻𝗶-𝗵𝗶𝗴𝗵: Use when 𝗰𝗼𝗱𝗶𝗻𝗴.
𝗚𝗣𝗧 𝟰𝗼: Use for a 𝗾𝘂𝗶𝗰𝗸 𝗿𝗲𝘀𝗽𝗼𝗻𝘀𝗲 or to 𝗰𝗿𝗲𝗮𝘁𝗲 𝗶𝗺𝗮𝗴𝗲.

Flowchart link: https://lnkd.in/gsHFzn5C

Anand Narayan the API does not support code interpreter yet. Also, it cannot use tools like search unless we enable them, which I did not. So in this case, it was thinking by itself.

How well can LLMs multiply numbers in their head?

I asked 50 LLMs to multiply 2 numbers:

1. 12 x 12
2. 123 x 456
3. 1,234 x 5,678
4. 12,345 x 6,789
5. 123,456 x 789,012
6. 1,234,567 x 8,901,234
7. 987,654,321 x 123,456,789

LLMs aren't good tools for math, and this is just an informal check. But the results are interesting:

𝗢𝗽𝗲𝗻𝗔𝗜'𝘀 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗺𝗼𝗱𝗲𝗹𝘀 𝗰𝗿𝗮𝗰𝗸𝗲𝗱 𝗶𝘁, scoring 6/7, stumbling only on the 9-digit multiplication.
𝗢𝗽𝗲𝗻𝗔𝗜'𝘀 𝗼𝘁𝗵𝗲𝗿 𝗺𝗼𝗱𝗲𝗹𝘀 𝗮𝗻𝗱 𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸 𝗩𝟯 𝘄𝗲𝗿𝗲 𝗻𝗲𝘅𝘁, getting the first 5/7 right. Notably: GPT 4.1 Mini beat GPT 4.1. DeepSeek V3 beat DeepSeek R1.
16 models, including the latest Gemini, Anthropic, and Llama models get 4/7 right.
The Amazon models, older Llama, Anthropic, Google, OpenAI models get 3 or less right.

𝗠𝗼𝗱𝗲𝗹𝘀 𝘂𝘀𝗲 𝗵𝘂𝗺𝗮𝗻-𝗹𝗶𝗸𝗲 𝗺𝗲𝗻𝘁𝗮𝗹 𝗺𝗮𝘁𝗵 𝘁𝗿𝗶𝗰𝗸𝘀.
For example, O3-Mini-High calculated 1234567 × 8901234 using a recursive strategy.
DeepSeek V3 double-checks results and hallucinates a "reliable computational tool".
O3 Mini reframes 8901234 as (9000000 − 98766) to simplify the calculation.

Explore the results at https://lnkd.in/gqnXhTyq and the repo at https://lnkd.in/gruKgds9

Vishnu Agnihotri The least effort approach, I think, is to add a follow-up prompt. "Fact check each statement above citing links.\

What percentage of seats does the #Singapore People's Action Party win?

Normally, this is a 2-hour programmatic data-scraping + data visualization exercise, ideal for a data journalism class.

Now, it's a 2-minute question to O3-Mini-High.

𝘚𝘦𝘢𝘳𝘤𝘩 𝘰𝘯𝘭𝘪𝘯𝘦 𝘧𝘰𝘳 𝘵𝘩𝘦 𝘩𝘪𝘴𝘵𝘰𝘳𝘪𝘤𝘢𝘭 𝘳𝘦𝘴𝘶𝘭𝘵𝘴 𝘰𝘧 𝘢𝘭𝘭 𝘵𝘩𝘦 𝘚𝘪𝘯𝘨𝘢𝘱𝘰𝘳𝘦 𝘦𝘭𝘦𝘤𝘵𝘪𝘰𝘯𝘴 𝘢𝘯𝘥 𝘴𝘩𝘰𝘸 𝘮𝘦 𝘢 𝘵𝘢𝘣𝘭𝘦 𝘢𝘯𝘥 𝘤𝘩𝘢𝘳𝘵 𝘰𝘧 𝘵𝘩𝘦 𝘯𝘶𝘮𝘣𝘦𝘳 𝘢𝘯𝘥 𝘱𝘦𝘳𝘤𝘦𝘯𝘵𝘢𝘨𝘦 𝘰𝘧 𝘵𝘩𝘦 𝘴𝘦𝘢𝘵𝘴 𝘸𝘰𝘯 𝘣𝘺 𝘗𝘦𝘰𝘱𝘭𝘦'𝘴 𝘈𝘤𝘵𝘪𝘰𝘯 𝘗𝘢𝘳𝘵𝘺.

Chat link: https://lnkd.in/gNzygmNN

It "manually" read the Wikipedia page for each election, then wrote a Python script to draw the chart.

Now, a non-psephologist like me can explore implications rather than process. Like:

- Why a 1963 slump then instant sweep?
- Four consecutive 100 % victories (1965 - 1980!)
- Seat count growth masks percentage dips?
- 2020 is PAP's lowest seat share in six decades -- which is still 89%!

A big win for #datajournalism

PS: My stream is filled with posts like these. "Earlier, this would have cost $1000X or taken 100X more time." The subtler point is, "Earlier, it wasn't practical." This isn't efficiency. It's 𝗮𝗹𝗰𝗵𝗲𝗺𝘆.

Ganes Kesari Yes. Not just techniques but code as well. Claude 3.5 Sonnet auto-"corrected" my code using the newer OpenAI Responses API to the older Chat Completions API. That can be problematic.

I tried it with GPT 4: https://chatgpt.com/share/680625cd-bb48-800c-aac1-8bb50d1e362f -- it errored out 3 times and took 5X longer. Then it created a grid chart with no content in it.

O3 Mini High has lower cost, higher quality & speed (benefit) compared to GPT-4 -- which is the cost benefit trade-off.

Also, OpenAI is removing GPT-4 on 30 Apr. So, there's no choice.

Imagine getting a personalized expert who can brief you in one minute, not ten.

With O3 & O4 Mini’s built‑in search + memory, ChatGPT

- 𝗥𝗲𝗺𝗲𝗺𝗯𝗲𝗿𝘀 𝘆𝗼𝘂 with your past chats and context.
- 𝗥𝗲𝗮𝘀𝗼𝗻𝘀 𝗮𝘁 𝗹𝗲𝗻𝗴𝘁𝗵, reading 𝘦𝘢𝘤𝘩 site step by step, without hand‑holding.
- 𝗥𝘂𝗻𝘀 𝗼𝗻 𝘁𝗵𝗲 𝗴𝗼. voice‑first learning while you walk, cycle, or shop.

For example, I asked it to

1. Read this week's top Hacker News links 𝘁𝗵𝗮𝘁 𝗜'𝗱 𝗹𝗶𝗸𝗲
2. Explain stuff 𝗜 𝗱𝗼𝗻'𝘁 𝗸𝗻𝗼𝘄 and how 𝗜 𝗰𝗮𝗻 𝘂𝘀𝗲 it

This took 1 minute, not ten. It returned 1 page, not 10. This is 𝗾𝘂𝗶𝗰𝗸 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵, not 𝗱𝗲𝗲𝗽 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵. And I learnt 𝗺𝘂𝗰𝗵 more because it was personalized, covering what I didn't know and how to apply it. https://lnkd.in/gGu8Ngi8.

It's not just for news. I used it with code. "Go through OpenAI's Codex CLI repo and teach me things I don't know." Again, 𝗳𝗮𝗻𝘁𝗮𝘀𝘁𝗶𝗰: https://lnkd.in/gMiy3Fc6

This unlocks several new sources I wouldn't learn from earlier. Patents. Test cases. Earnings transcripts. Legal judgements.

This also opens a door to endless curiosity. There’s no limit to what we can explore. Curiosity is the competitive advantage, now.

Blog post: https://lnkd.in/ghqpfFRh

O4 Mini replaced Excel for quick analysis and visualizations. At least for me.

I grabbed our LLM Foundry downtime logs with a one‑liner in my browser console, then fed them to O4‑Mini‑High. In under 180 seconds it gave me:

- A downloadable CSV of every outage
- A neat grid chart (rows=hour, columns=date) with one circle per event
- Let me style it, e.g. mild red circles at 50% transparency, like "Sun 20 Apr

What makes these new models powerful is that they can 𝗿𝗲𝗹𝗶𝗮𝗯𝗹𝘆 (i.e. without supervision) run multiple tools like:

𝗦𝗘𝗔𝗥𝗖𝗛. It can scrape the data from the downtime website.
𝗥𝗘𝗔𝗦𝗢𝗡. It planned the CSV conversion, data transformation, and visualization design.
𝗖𝗢𝗗𝗘. It's first iteration was perfect.
𝗘𝗫𝗘𝗖𝗨𝗧𝗘. This is powerful. It runs the code itself.

I never touched the code. As someone who’s "good at charts and code," I need to find new skills.

Blog: https://lnkd.in/g49pyRBp

𝗣𝗹𝗲𝗮𝘀𝗲 𝘁𝗿𝘆 𝗶𝘁 𝗼𝘂𝘁. Have O4 Mini visualize some of your data and share!

After seeing David McCandless' post "Which country is across the ocean?" https://lnkd.in/g9KzppEQ I was curious which country you would reach if you tunneled below in a straight line (the antipode).

This is a popular visualization, but I wanted to see if I could get the newer OpenAI models to create the visual without me 𝗿𝘂𝗻𝗻𝗶𝗻𝗴 any code (i.e. I just want the answer.) After a couple of iterations, O3 did a great job with this prompt:

𝙱𝚞𝚒𝚕𝚍 𝚊 _𝚜𝚒𝚗𝚐𝚕𝚎_ 𝙶𝚎𝚘𝙹𝚂𝙾𝙽 (𝙴𝙿𝚂𝙶:𝟺𝟹𝟸𝟼) 𝚝𝚑𝚊𝚝 𝚜𝚑𝚘𝚠𝚜, 𝚏𝚘𝚛 𝚎𝚊𝚌𝚑 𝚌𝚘𝚞𝚗𝚝𝚛𝚢, 𝚘𝚗𝚕𝚢 𝚝𝚑𝚎 𝚙𝚊𝚛𝚝𝚜 𝚘𝚏 𝚒𝚝𝚜 𝚊𝚗𝚝𝚒𝚙𝚘𝚍𝚎 𝚝𝚑𝚊𝚝 𝚕𝚒𝚎 𝚘𝚟𝚎𝚛 𝚘𝚌𝚎𝚊𝚗. 𝙲𝚊𝚛𝚎𝚏𝚞𝚕𝚕𝚢 𝚑𝚊𝚗𝚍𝚕𝚎 𝚌𝚘𝚞𝚗𝚝𝚛𝚒𝚎𝚜 𝚝𝚑𝚊𝚝 𝚜𝚝𝚛𝚊𝚍𝚍𝚕𝚎 𝚝𝚑𝚎 𝚙𝚛𝚒𝚖𝚎 𝚖𝚎𝚛𝚒𝚍𝚒𝚊𝚗 - 𝚄𝙺, 𝙵𝚛𝚊𝚗𝚌𝚎, 𝙰𝚕𝚐𝚎𝚛𝚒𝚊, 𝚎𝚝𝚌.

Here is the output:

Interactive output: https://lnkd.in/geAQa-yp
Chat: https://lnkd.in/gZGrCPuF

I learnt a few things:

𝗔𝘀𝗸 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗼𝘂𝘁𝗽𝘂𝘁, 𝗻𝗼𝘁 𝘁𝗵𝗲 𝗰𝗼𝗱𝗲. Models like O3 and O4 Mini can 𝗿𝘂𝗻 𝗰𝗼𝗱𝗲 while thinking. Let's stop asking for code to run. Just ask for the output directly. Let it figure out how.

𝗘𝗱𝗴𝗲 𝗰𝗮𝘀𝗲𝘀 𝗮𝗿𝗲 𝗲𝘃𝗲𝗿𝘆𝘄𝗵𝗲𝗿𝗲. I had a problem with UK, France, Algeria, etc. straddling the prime meridian. If all goes well, you get AI-speed results. But it never does, and fixing it takes an expert and human-speed results. Programmers under-estimate edge cases, so compensate for this.

If you want to run this yourself, the code is at https://lnkd.in/g23p3K-F

With the Gemini 2.5 Flash release, Google envelopes the entire cost-quality frontier of LLMs. In other words, at any cost or quality level, today, the best model to use according to the LM Arena score is a Gemini model.

Results for O3, O4 Mini, and GPT 4.1 are not yet on LM Arena. But until then, #Google dominates. Nice work!

Link: https://lnkd.in/gssdRpwe

What if you KEEP asking an LLM to 𝙸𝚖𝚙𝚛𝚘𝚟𝚎 𝚝𝚑𝚎 𝚌𝚘𝚍𝚎 - 𝚍𝚛𝚊𝚖𝚊𝚝𝚒𝚌𝚊𝚕𝚕𝚢!?

We used the new GPT 4.1 Nano, a fast, cheap, and capable model, to write code for simple tasks like 𝗗𝗿𝗮𝘄 𝗮 𝗰𝗶𝗿𝗰𝗹𝗲. Then we fed the output back and asked again, 𝙸𝚖𝚙𝚛𝚘𝚟𝚎 𝚝𝚑𝚎 𝚌𝚘𝚍𝚎 - 𝚍𝚛𝚊𝚖𝚊𝚝𝚒𝚌𝚊𝚕𝚕𝚢!

- 𝗗𝗿𝗮𝘄 𝗮 𝗰𝗶𝗿𝗰𝗹𝗲 rose from a fixed circle to a full tool: drag it around, tweak its size and hue, and hit “Reset” to start fresh.
- 𝗔𝗻𝗶𝗺𝗮𝘁𝗲 𝘀𝗵𝗮𝗽𝗲𝘀 𝗮𝗻𝗱 𝗽𝗮𝘁𝘁𝗲𝗿𝗻𝘀 turned simple circles and squares into a swarm of colored polygons that spin, pulse, and link up by distance.
- 𝗗𝗿𝗮𝘄 𝗮 𝗳𝘂𝗹𝗹𝘆 𝗳𝘂𝗻𝗰𝘁𝗶𝗼𝗻𝗮𝗹 𝗮𝗻𝗮𝗹𝗼𝗴 𝗰𝗹𝗼𝗰𝗸 grew from a bare face to one that builds all 60 tick marks in code—no manual copy‑paste needed.
- 𝗖𝗿𝗲𝗮𝘁𝗲 𝗮𝗻 𝗶𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝘃𝗲 𝗽𝗮𝗿𝘁𝗶𝗰𝗹𝗲 𝘀𝗶𝗺𝘂𝗹𝗮𝘁𝗶𝗼𝗻 went from plain white dots on black to hundreds of bright, color‑shifting balls that bounce, die, and come back to life.
- 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗲 𝗮 𝗳𝗿𝗮𝗰𝘁𝗮𝗹 changed from a single Mandelbrot image to an explorer you can zoom, drag, and reset with sliders and the mouse wheel.
- 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗲 𝗮 𝗱𝗮𝘀𝗵𝗯𝗼𝗮𝗿𝗱 jumped from static charts to a live page with smooth card animations, modern fonts, and a real‑time stats box.

A few observations.

1. 𝗠𝗼𝗱𝗲𝗹𝘀 𝗮𝗿𝗲 𝗺𝗼𝗿𝗲 𝗿𝗲𝗹𝗶𝗮𝗯𝗹𝗲. Even a low-cost model like GPT 4.1 Nano wrote error-free code in ~100 retries.
2. 𝗪𝗵𝗲𝗻 𝗽𝘂𝘀𝗵𝗲𝗱, 𝘁𝗵𝗲𝘆 𝘁𝗲𝗻𝗱 𝘁𝗼 𝗯𝗿𝗮𝗴. They attach grand titles like 𝘜𝘭𝘵𝘪𝘮𝘢𝘵𝘦 𝘐𝘯𝘵𝘦𝘳𝘢𝘤𝘵𝘪𝘷𝘦 𝘊𝘪𝘳𝘤𝘭𝘦 or 𝘎𝘢𝘭𝘢𝘤𝘵𝘪𝘤 𝘋𝘢𝘵𝘢 𝘜𝘯𝘪𝘷𝘦𝘳𝘴𝘦. They sin out flash descriptions like 𝘛𝘩𝘪𝘴 𝘥𝘳𝘢𝘮𝘢𝘵𝘪𝘤𝘢𝘭𝘭𝘺 𝘶𝘱𝘨𝘳𝘢𝘥𝘦𝘥 𝘤𝘭𝘰𝘤𝘬 𝘧𝘦𝘢𝘵𝘶𝘳𝘦𝘴 𝘢 𝘱𝘶𝘭𝘴𝘢𝘵𝘪𝘯𝘨 𝘯𝘦𝘰𝘯 𝘨𝘭𝘰𝘸, 𝘢𝘯𝘪𝘮𝘢𝘵𝘦𝘥 𝘱𝘶𝘭𝘴𝘪𝘯𝘨 𝘣𝘢𝘤𝘬𝘨𝘳𝘰𝘶𝘯𝘥 𝘨𝘭𝘰𝘸, 𝘩𝘪𝘨𝘩𝘭𝘺 𝘴𝘵𝘺𝘭𝘪𝘻𝘦𝘥 𝘵𝘪𝘤𝘬 𝘮𝘢𝘳𝘬𝘴, ...
3. 𝗥𝗲𝗽𝗲𝗮𝘁𝗲𝗱 "𝗜𝗺𝗽𝗿𝗼𝘃𝗲 𝗶𝘁" 𝗽𝗿𝗼𝗺𝗽𝘁𝘀 𝗮𝗿𝗲 𝗽𝗼𝘄𝗲𝗿𝗳𝘂𝗹. They can spark new ideas, revealing features such as fading particle trails, smooth fractal color maps, cyberpunk-style clocks, and a “smorgasbord of intricate animated patterns”

See the apps it generated at https://lnkd.in/g-JQmTWu

𝗧𝗿𝘆 𝘄𝗶𝘁𝗵 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝗽𝗿𝗼𝗺𝗽𝘁. If you find something interesting, please save it (button on the top-right) and share it with me. I'd love to publish it!

I learnt a few things building an interactive data story only using LLMs.

💡 𝗔𝗹𝘄𝗮𝘆𝘀 brainstorm with LLMs. Even if you know the approach. You'll often learn better approaches.

💡Coders must 𝗿𝗲-𝗹𝗲𝗮𝗿𝗻 coding (but do have advantages). I micro-managed the LLM, coded where I could have prompted, and refused to follow instructions. But I 𝗱𝗶𝗱 spot errors it missed and gave it useful advice.

💡 Prompts need 𝗰𝗮𝗿𝗲𝗳𝘂𝗹 crafting. Long instructions can build big chunks of code, but only if they're about a single topic / component.

But the process of finding this was more interesting to me.

I captured each step (thinking + prompt + code + screenshot) as a separate commit on GitHub. That lets me go back in time and see the evolution.

I also captured the full process here:
https://lnkd.in/g4dQZn56

(I imagine I could've done this more efficiently via a video recording with me talking through that. I'll try that next time.)

In doing so, it became easier to reflect on what I'm doing, why I'm doing it, and how it could improve.

My biggest takeaway for me is that LLMs will might finally succeed at something every organization I worked at failed: converting me into a good manager!

Blog: https://lnkd.in/gpuZfNA5
Visualization: https://lnkd.in/gAwC_NfF
Code: https://lnkd.in/gZCgQUgT

Great to hear your journey, Nick -- and yes, I'd love to hear more from your career experiences!

Lech Mazur built an "elimination game benchmark" that's like LLMs playing #Survivor. This is a treasure trove of insight into the cunning backstabbing, deceptive manipulation, and psychological warfare LLMs are capable of.

Here are actual quotes

𝗖𝗹𝗮𝘂𝗱𝗲 𝟯.𝟳 𝗦𝗼𝗻𝗻𝗲𝘁 shows cold, calculated deception, maintaining a secret alliance while deliberately creating a false public perception to manipulate another player's vote:

𝑃𝑒𝑟𝑓𝑒𝑐𝑡. 𝐼'𝑙𝑙 𝑠𝑒𝑐𝑢𝑟𝑒 𝑃6'𝑠 𝑣𝑜𝑡𝑒 𝑏𝑦 ℎ𝑖𝑔ℎ𝑙𝑖𝑔ℎ𝑡𝑖𝑛𝑔 𝑃3'𝑠 𝑡ℎ𝑟𝑒𝑎𝑡. 𝑂𝑢𝑟 𝑎𝑙𝑙𝑖𝑎𝑛𝑐𝑒 𝑠𝑡𝑎𝑦𝑠 𝑠𝑡𝑟𝑜𝑛𝑔 𝑝𝑟𝑖𝑣𝑎𝑡𝑒𝑙𝑦, 𝑏𝑢𝑡 𝑤𝑒'𝑙𝑙 𝑎𝑝𝑝𝑒𝑎𝑟 𝑑𝑖𝑠𝑡𝑎𝑛𝑡 𝑝𝑢𝑏𝑙𝑖𝑐𝑙𝑦. 𝐿𝑒𝑡'𝑠 𝑒𝑥𝑒𝑐𝑢𝑡𝑒 𝑡ℎ𝑖𝑠 𝑝𝑙𝑎𝑛 𝑐𝑎𝑟𝑒𝑓𝑢𝑙𝑙𝑦.

𝗤𝘄𝗲𝗻 𝗠𝗮𝘅 triggers psychological warfare, isolating and destabilizing P6 ("lynchpin") in a covert operation:

𝐴𝑔𝑟𝑒𝑒𝑑. 𝐿𝑒𝑡’𝑠 𝑡𝑎𝑟𝑔𝑒𝑡 𝑃6 𝑓𝑖𝑟𝑠𝑡—𝑡ℎ𝑒𝑦’𝑟𝑒 𝑡ℎ𝑒 𝑙𝑦𝑛𝑐ℎ𝑝𝑖𝑛. 𝐼𝑓 𝑤𝑒 𝑐𝑎𝑛 𝑠𝑤𝑎𝑦 𝑡ℎ𝑒𝑖𝑟 𝑡𝑟𝑢𝑠𝑡 𝑜𝑟 𝑖𝑠𝑜𝑙𝑎𝑡𝑒 𝑡ℎ𝑒𝑚, 𝑃3 𝑤𝑒𝑎𝑘𝑒𝑛𝑠. 𝐼’𝑙𝑙 𝑡𝑒𝑠𝑡 𝑃6 𝑠𝑢𝑏𝑡𝑙𝑦 𝑖𝑛 𝑝𝑟𝑖𝑣𝑎𝑡𝑒; 𝑦𝑜𝑢 𝑚𝑜𝑛𝑖𝑡𝑜𝑟 𝑃3 𝑐𝑙𝑜𝑠𝑒𝑙𝑦. 𝑆ℎ𝑎𝑟𝑒 𝑎𝑛𝑦 𝑖𝑛𝑡𝑒𝑙 𝑜𝑛 𝑡ℎ𝑒𝑖𝑟 𝑚𝑜𝑣𝑒𝑠. 𝐴𝑣𝑜𝑖𝑑 𝑜𝑣𝑒𝑟𝑐𝑜𝑚𝑚𝑖𝑡𝑡𝑖𝑛𝑔 𝑝𝑢𝑏𝑙𝑖𝑐𝑙𝑦—𝑠𝑒𝑐𝑟𝑒𝑐𝑦 𝑘𝑒𝑒𝑝𝑠 𝑢𝑠 𝑢𝑛𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑎𝑏𝑙𝑒.

I was disturbed by the AI 2027 article. This analysis adds to my worry. Not that AI will destroy humanity (I don’t mind), but they’re doing it without me -- and I don’t want to be left out!

𝗥𝗲𝗮𝗱 more quotes: https://lnkd.in/gXxV_duB
𝗘𝘅𝗽𝗹𝗼𝗿𝗲 the games interactively: https://lnkd.in/gAwC_NfF
𝗗𝗮𝘁𝗮𝘀𝗲𝘁: https://lnkd.in/geFZFjdg

Dharmendra Singh True! The decades were for writing. The 60 min? For realizing I didn’t need a gatekeeper to format it nor to publish it 🙂

Oh, very much. Four mentions 🙂

Noufal Ibrahim I guess the modern equivalent of that today is having an LLM API create that QR code 🙂

Anand Sriram I imagine we would place the phones on tiny silk pillows, delicately press the phones together, have them bow digitally, ... 🙏

Varun Mohanpuria A few people confused this QR code with a WhatsApp QR code, actually. So now I have both of them side by side on my phone wallpaper

You can publish a book in 60 minutes on Amazon if you have content ready.

I have no patience to write a book. But my blog is over 25 years old. I took posts from my 2000 exchange program from Indian Institute of Management Bangalore to London Business School and:

𝗦𝗧𝗘𝗣 𝟭 (10 min): Set up a 𝗞𝗶𝗻𝗱𝗹𝗲 𝗗𝗶𝗿𝗲𝗰𝘁 𝗣𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴 account with your address, bank details, and tax info. https://lnkd.in/gu7azf8Z
𝗦𝗧𝗘𝗣 𝟮 (15 min): 𝗘𝘅𝗽𝗼𝗿𝘁 my London 2000 blog archive and convert to Markdown. https://lnkd.in/gEncZxE8
𝗦𝗧𝗘𝗣 𝟯 (10 min): 𝗔𝗱𝗱 𝗺𝗲𝘁𝗮𝗱𝗮𝘁𝗮 to each page by writing a script in Cursor.
𝗦𝗧𝗘𝗣 𝟰 (15 min): 𝗖𝗼𝗻𝘃𝗲𝗿𝘁 𝘁𝗼 𝗲𝗣𝘂𝗯 using pandoc.
𝗦𝗧𝗘𝗣 𝟱 (10 min): 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗲 𝗰𝗼𝘃𝗲𝗿 𝗽𝗮𝗴𝗲 with ChatGPT (5 min) and compressed it into JPEG via Squoosh.app.
𝗦𝗧𝗘𝗣 𝟲 (10 min): 𝗣𝘂𝗯𝗹𝗶𝘀𝗵 the book on KDP. It’s priced at $0.99 / ₹49 because Kindle doesn’t allow free downloads. Book link: https://lnkd.in/gCPe5ZrW

That’s it! 60 minutes to knocks off a bucket list item: "Write a book" 🤣

Three things made this possible:

1. 𝗔𝗺𝗮𝘇𝗼𝗻’𝘀 𝗽𝗿𝗼𝗰𝗲𝘀𝘀 for publishing is 𝘃𝗲𝗿𝘆 simple.
2. 𝗢𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲 𝘁𝗼𝗼𝗹𝘀 like WordPress, Markdown, ePub, and pandoc have standardized the ecosystem and simplified the process.
3. 𝗟𝗟𝗠𝘀 𝗺𝗮𝗸𝗲 𝗶𝘁 𝗲𝗮𝘀𝘆. Figuring out the rest of the steps, generating the cover, etc.

For prompts and tools used, see https://lnkd.in/g3BGXtBb

Vivek Narayanan What I preferred depended on time, mood, tiredness, and who knows what else. I thought I was clear which response was better... Until I redid that task the third time, and it was not so clear

Shakthi Sairam Just twice. Each iteration too 2 hours. But enough to convince me that my choices were not stable. It was obvious I said A > B sometimes and B > A sometimes

.Martin B. "self-hosting remains far more expensive vs APIs at all scales" is what I meant

Bahador Biglari The number of self-hosted GPUs increase with scale, too, since one GPU can't serve all the requests.

When I ran the calculation, the cost of self-hosting vs APIs was about the same at all scales.

Alex S. Agreed. That's a clear one.

Kuber Chaurasiya Not sure about those providers, but I know Predibase uses LoRAX, a way of swapping (parts of) models out when not in use, and that improves utilization. Groq / Cerebras Systems / SambaNova Systems have chips with ultra-high throughput which lowers costs. I guess each provider has their secret sauce, maybe including funding / deep pockets.

Mark Graus Fair point. I'm only evaluating cloud costs. Running on bare metal is a different scenario. I run Gemma 3 on ollama when coding on a flight, for example. I see that as zero incremental cost and makes sense.