DeepSeek Shows Humility Where Sam Altman Cannot

Author

Neer Varshney
March 25, 2025

In This Newsletter

→ DeepSeek’s killer improvements to V3

→ Perplexity CEO says y’all better prepare to lose your jobs

→ Paul Graham and ChatGPT writing like kids on assignments

→ Writing better prompts

→ Trump v/s Zelenskyy moment of Cloud Computing and other AI happenings

→ Be Featured on Artificially Boosted Wall of Love

When mainstream media hyped up paranoia around DeepSeek, it mostly touted the geopolitical battle between the U.S. and China.

I remain fairly pissed to be honest, because there were so many interesting discussions to be had there, only if we had some scope for nuance.

Anyway, while all of the doomsayers have moved on from talking about DeepSeek, ‘the Chinese company’ keeps delivering.

DeepSeek on Monday dropped what it said was a “small update” to its V3 model — only that the small update pushed it toward the top of benchmarks for best coding models available to the public right now.

talk show the game show it wasnt me GIF by truTV

Reliable benchmark analyst kcores-llm-arena was among the first to analyze the model and published results that show the model to now be placed at No. 3 for coding performance, next to Claude 3.7 Sonnet Thinking and Claude 3.5 Sonnet.

There are a lot of variables when determining “best LLM” for any given use-case but the benchmarks do give a good indication. And kcores’ results are in line with my personal experience.

I just highlighted a couple of newsletters ago how Claude series is my go-to choice for coding and how Claude 3.5 Sonnet is performing better than the vanilla 3.7 version for me.

On OpenAI’s end, o1 is undoubtedly the best coding model I have used from the company. But I simply don’t use it that frequently and the reason is very simple.

Imagine yourself in my shoes too with the table below. If you could use your favorite coding model at $15 output or DeepSeek's crazy good alternative at around a dollar or two — why would you care for a $60 model, even if it is still super good?

Type

Claude 3.7 / 3.5

DeepSeek V3

OpenAI o1

Gemini 2.0 Pro

Input

$3

$0.135 to $0.27

$15

$0

Output

$15

$0.135 to $2.19

$60

$0

All pricing are per million token and cache-based pricing is ignored, but they are always proportional.

But economic pragmatism aside, I can’t help but see this in another light of the philosophical debate around the whole U.S. and China thing.

That there is an entirely different approach to building in the east and west and it would become more apparent with more global companies coming out China and India.

American companies are deeply bullish by nature, and that is possibly what leads Sam Altman to build the “AGI is here” narrative with every new launch, often disappointing even die-hard OpenAI fans.

Meanwhile, DeepSeek just casually drops new reasoning models at a fraction of the cost or optimizes its models to be best at coding, without much furor.

Perplexity CEO On AI Eating Up Jobs

If you all missed it, Aravind Srinivas went to Indian billionaire Nikhil Kamath’s podcast, and said what we all should know by now — that AI is already displacing jobs across the web economy.

Big Tech is laying off and not hiring for any new roles on a net basis. A lot of roles posted on platforms like LinkedIn are simply ghost postings.

Whether you are a developer, copywriter, journalist, graphics designer — be prepared, companies simply don’t need that much work done from humans anymore.

Does that mean jobs across the web economy are going to disappear? Absolutely not, but one highly-skilled professional armed with AI skills can do what took a town to achieve before.

Paul Graham articulated what I couldn’t

Among the biggest frustrations of conversing with AI is it not knowing the intent and flow of the conversation — it’s a side-effect of the token-based prediction models that we have right now.

Paul Graham said it perfectly. AI needs to get better at the “emotional” quotient and stop behaving like kids writing their assignments. I think OpenAI took a step in the right direction with GPT-4.5-Preview here and this is a sphere where Grok has also been making progress.

Builders’ Corner

This is the section where we explore building stuff with AI for people that are not experienced coders

I wanted to focus a bit on “prompt engineering” for our builders today. Unlike what the word implies, there are no weird greek letters involved here.

Many treat AI like a magic 8-ball: toss a half-formed idea at ChatGPT, expect perfection, then give up when it fails.

But not my readers. They treat AI like a junior teammate. You wouldn’t tell a human, “Do something cool,” and expect great results. Here’s how to refine your prompts:

🚫 Bad Prompt Example 1: Too Vague

Prompt: “Improve my email.”

Why it fails: No context, no goal. Is this a sales pitch? A apology? A newsletter? The AI flails in the dark.

Improved Prompt:

Rewrite this email to persuade a busy client to extend our project deadline by 1 week. Friendly but professional tone.

- Not more than 150 words

- i really wanna emphasize that the extension will ensure higher quality results even while being subtly acknowledging their time constraints

- offer discount on next project without being unnecessarily cocky

The AI now has structure, intent, and guardrails to work with.

Result: 🚫 Bad Prompt Example 2: No Scope or Direction

Prompt: “Generate a website about a snake game for me.”

Why it fails: What kind of website? A landing page? A tutorial? A portfolio? The game itself? The AI will guess (poorly). Yeah one of the readers did do this last time.

Improved Prompt

Create a playable snake game for me using html/css/js, feel free to use any framework like bootstrap or nippleJS or threeJS as you prefer

Let’s make sure, it is:

- mobile friendly

- classic arcade feel

- has high scores, turbo speed boosts, three difficulty modes

The AI delivers a focused, functional template instead of generic fluff.

🚫 Bad Prompt Example 3: Missing Audience Context

Prompt: “Write a workout plan.”

Why it fails: For who? A marathon runner? A post-injury rehab? A busy parent?

Improved Prompt

Design a 4-week home workout plan for a beginner 35-yr old female with no equipment.

Goal: build strength and stamina.

-3 sessions/week (45 mins max)

-Warm-up/cool-down routines

-Modifications for knee arthritis

-Focus on bodyweight exercises like squats, planks, etc.

-i hate cardio

Tailored, actionable, and avoids irrelevant advice (like heavy lifting).

🛠️ Pro Tip: The 3-Step Prompt Framework

Before asking AI, define:

  1. Who: Audience/end-user (e.g., “busy clients,” “retro gamers”).

  2. What: Goal/action (e.g., “persuade,” “promote,” “teach”).

  3. How: Style/constraints (e.g., tone, format, word count, tools).

The Takeaway: Specificity = Success. The more you mimic how you’d explain a task to a human (details, examples, boundaries), the better AI performs.

Don’t bother too much with spellings or even mixing up languages if you are bilingual like me. AI is good at that. Just ensure that it has enough knowledge and logical sequence to work with.

Struggling with a prompt? Reach out — I’ll help you engineer it! 💪

Latest Happenings

  • OpenAI announced some leadership responsibilities shuffle with COO Brad Lightcap increasing focus on “day-to-day“ business operations, freeing up more of Sam’s time to be spent on tech and research front, as CTO role remains unfilled after Mira Murati’s depature in September.

  • In a Trump v/s Zelenskyy moment of cloud computing, Vercel and Cloudflare execs are taking public shots at each other after Vercel fixed a security vulnerability for next.js users on its platform but kinda kept the whole thing under the rug for the users of the open-source framework elsewhere.

  • Cloudflare is using generative AI to lure mischievous web scrappers into gibberish text to misguide them away from actual site content. I think would be interesting to see if this adds enough costs for scrappers to make using LLM as a solution to pass this “labyrinth.”

  • Launching a reasoning model has pretty much become a must for every LLM out there now and Tencent is latest to join the party with its T1.

From Artificially Boosted Family

I started this newsletter to answer AI questions from friends and family, and didn’t expect to become such a thing!

Now that it has, we have rolled out a landing page https://artificiallyboosted.com/.

To be featured on the Wall of Love, simply post a review of the newsletter on any social media of your choice and tag me on that platform! Or respond to this email!

I put some genuine effort into writing this from a point of view of educating my readers alongside me, so I appreciate if you can spare a minute to throw some feedback at me.

Meanwhile, don’t forget to check out Dzambhala Nodex if you invest or trade in the US or India markets!

Reply

or to participate.