This post was originally published on this site.
OpenAI’s DeepResearch, Replit’s App Developer App, Google’s Gemini 2 — the end of the low-level jobs is here.
Our day-to-day world is filled with inefficiencies. Capitalism works because it allows anyone to fend for those inefficiencies by developing some skills in return for monetary profits.
With time, some of these skills end up at the lower end of value creation.
This deprecation of skills happens through either technological innovation (called Technological Unemployment) or through the dark hand of economics – the market adjusting to the disbalance of supply/demand
It was this economic disbalance that became a boon for Indians in the 1990s and made India the dumping ground for low-end jobs of the West. It did bring us some equitable growth. Think BPO, tech consulting, call center operative, analysis, operational work, etc.
But, times have changed and this time, it’s the bane of deprecation through technical innovation that’s upon us (and everywhere else for that matter).
This chart compares the accuracy of an LLM model’s responses to the GPQA Dimond test over time. (This test contains tough 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. These are hard questions. So much so that PhD students, using Google, can only reach ~65% accuracy 😵💫)
Hence, OpenAI’s o1-pro and recently launched o3 LLM models have already overtaken the accuracy (and hence reasoning capability) of human PhDs.
So, we’re at a stage where the best model till the end of last week was already smarter, and MUCH cheaper than any human on the earth.
What does that mean for you?
This is the distribution of jobs in the US from the 1950s to 2017 (each section determines the jobs in the sector)
If you look at Sales and Administrative Support the % of the population employed is almost the same as they were in the 1950s — guess where I believe the next contraction due to AI will be?
A few entry-level jobs were replaced by BigAI last week, here’s the run-down:
-
👑 Google 👑 makes a Comeback with Agentic Gemini 2.0
Often the butt of AI jokes, Google launched Gemini 2 — their most powerful, agentic, multi-modal, low-latency, high-performance, and CHEAP family of LLMs and well, for the first time it seems that they delivered!
They launched an entire family of Gemini 2 models together, each catering to a specific need. Here’s the brief on each of them:
-
Gemini 2.0 Flash: This model is like the Usain Bolt of AI, doubling the speed of its predecessor, Gemini 1.5 Pro. Built for agentic experience with better speed, better performance, and multi-modal i.e. can handle native image generation and multilingual text-to-speech.
-
Gemini 2.0 Flash-Lite: The budget-friendly hero for everyone who wants to dip their toes into AI without breaking the bank. Perfect for tasks where you need massive text output with quality. Think class project, product research, etc.
-
Gemini 2.0 Pro: A multi-modal powerhouse of an LLM. With a context window of 2 million tokens, this model can handle anything from coding challenges to deep dives into world knowledge.
-
Gemini 2.0 Flash Thinking (Experimental): The reasoning-specific model from the family. Think DeepSeek R1 from Google. But, this comes with a bigger context window of 1 million tokens (ChatGPT’s still locked at 128k) that supports code execution too. WOW!
And all this kills at the benchmarks!
Very hard to distinguish between each of them right?! They’re suffering from success it seems.
But, why should you care?Google’s dignified launch of Gemini 2.0 shows that the company has some zeal left in it — something that was missing in all its prior releases. Being the creator of the Transformers architecture and being the poster boy for AI/ML breakthroughs in SV, Google’s finally coming out as a strong AI competitor to OpenAI with this release.
In addition to this, Google is stepping into a world of AI agents. Projects like Astra, an assistant on your phone and glasses; and Mariner, an AI agent on the web to automate your online tasks — are a few of the projects that are being powered through Gemini 2.0
Google is like:
You can build your agents, train your RAG model, or fine-tune your models through Google AI Studio. All of this at the scale and price point that Gemini APIs are at — the agents very well could be built on top of the Google stack!
GET BUSY BUILDINGGGG Y’ALL!!!
-
-
OpenAI came for Research Analysts and — ATE!!!
OpenAI launched DeepResearch — a research agent and broke the scales, previously broken by their models, on Humanity’s Last Exam — a set of 3,000 challenging questions across over a hundred subjects ranging from chemistry, maths, physics, ecology, etc.
This agent is built on top of their upcoming o3 model. They’ve only publicly released its variations o3-mini and o3-mini-high which themselves have surpassed a lot of benchmarks — including reining Humanity’s last exam before this.
Now, this is good stuff — breaking benchmarks week after week is not an easy feat. Kudos to the OpenAI team. But, a $200/month plan, only 100 queries per month, and a context window of only 128K tokens — wasn’t appealing to a lot of folks.
Now, what DeepResearch is doing isn’t ground-breaking. It’s been done by Google first — isn’t everything? 😭. It’s just an agent with advanced reasoning and web-browsing capability. So, technically, anyone can build it with any LLM — so they did.
Few good contenders emerged:
-
Deep Research — fully open-source, utilizes OpenAI’s API. So you get an OpenAI model doing what DeepResearch does, without paying $200
-
and many more unique implementations.
The point is, with enterprise and open-source tools, the role of the mid-level consultant/analyst just evaporated.
-
-
Replit took an attempt at Mobile App Developers 📱🤖
Replit is an online platform that allows you to create apps using their platform.This week they launched their mobile app that allows you to build apps via their apps. You heard it right, an app that makes an app.
The app consists of an AI agent that takes in a prompt, suggests a few additional steps that it recognizes from the prompt, helps you visualize the output, tests it, and deploys the same.
The product isn’t in its best form yet. In our testing, we found it to be bugging now and then, but this does push the company nearer to the goal.
It’s just another Sonnet 3.5 wrapper, why should I care?
The idea of 1 person companies is picking steam on the internet. Popularized by the famous @levelsio — as building tech products gets commoditized, the future does seem to belong to the people with ideas. The true value would be derived from the creativity and taste of the apps/solutions rather than the one-size-fits-all apps. Think burger – McD, BurgerKing, and Wendy’s — all making the same thing differentiated by their marketing. And in such a market, a product like Replit, on mobile, really helps.
I can imagine the next big app built by an influencer with no coding experience.
Mobile developers for one second were like
-
Europe’s only AI-redeemer play 🇫🇷 🤖
French AI Lab, Mistral, launched their updated LeChat on both iOS and Android.
LeChat is a conversational AI (like ChatGPT). It’s multi-lingual and multimodel (unlike ChatGPT). Built on a MoE architecture. Is friggin fast. And, unlike Gemini, depending on the query, it redirects your query to their underlying models — Mistral Large, Mistral Small, and a prototype model called Mistral Next.
Anything special?
-
Speedy Responses: LeChat doesn’t just chat; it zips through conversations at the speed of light, offering up to 1,100 tokens per second. That’s faster than you can say “croissant”!
-
Canvas: With LeChat’s canvas feature, you can edit and transform your ideas in real time. It’s like a digital sketchbook where your thoughts come to life!
See both of these working together here:
And the best part, Mistral heavily embraces open-source culture. Not only do they make their technologies freely available, but all their Large Language Models (LLMs) are accessible on HuggingFace.
Vive LeChat!
Anyway, that’s it from our end. I built a Pomodoro timer with LeChat (inspired by this post) and Cursor last night and I’m psyched to build more this weekend!!!!!
Anyway, hope you have a great weekend ahead.
Until next Saturday!
PS: Let me know if this was succinct, you can shoot away your AI-related questions or reviews here: This Week In AI: Review, feedback, and query form
-