World in 2025
Humans have enjoyed ruling the Earth for 300,000 years. However,
for the first time since the Anthropocene epoch, humans have begun
to worry about the rise of a new species: AI.
Although AI has existed for over 60 years, the widespread concern
began with the release of ChatGPT in December 2022. That moment
marked a turning point, changing the world forever.
Initial frontier to Gen AI
Currently there are 6 frontier models in the limelight. 1.GPT
2.Deepseek 3.Claude 4.Grok 5.Llama 6.Gemini
ChatGPT initiated the race to AGI with GPT-3. The
closed source models GPT, Gemini, Claude & Grok were enjoying the
American VC money untill the launch of deepseek-r1 on Trump's
inauguration day. Even though Deepseek has released a bunch of
other models before R1 this caught attention mainly because of the
claimed model training cost. Only $5million. This was the time
when Trump already announed $500 billion stargate project to
domniate AGI. This broken US stock market especially Nvidia.
Open source models trend went from Falcon --> Mistal --> Llama -->
Deepseek. Now Mistral made a good comeback after the launch of Le
Chat, their mobile app which has the highest inference speed for
any LLM. They can answer upto 1000 words per second. Mistral is
using chips from Cerebras, an American company which specifically
builds chips for inference. Alibaba's Qwen is also another good OS
model to look out for.
Zuck announced that Meta is
targeting for SOTA Llama4, AI SWE to contribute to RnD, to power
this 2+ GW datacenter at Manhatten, bring online ~1GW power of
compute and targeting 1.3 million GPUs in 2025.
Bank of China is planning to provice 1 trillion yuan ($137
billion) in next 5 years to counter US's project Stargate.
Image Gen
DallE, Midjourney, Stable Diffusion, FLUX are the mojor terms in
image generation. Grok, Gemini, Llama, Mistral can generate
images. Currently Anthropic and Deepseek haven't relesed models
with image gen capability.
Video Gen
The tech world went berserk after seeing the demo of Sora in
February 2024. However, as of December 2024, only the demo exists,
and nothing has been officially released. Now, we have a lot of
video-generation AIs, such as Kling (China), Runway ML, Pika Labs,
and Luma (US).
The impressive part is that open-source models like Mochi1 (US)
and Hunyuan (China) are on par with the closed ones. Pepsi has
already released an ad fully generated by AI, and I’ve created
some good 30-second animated short films. I’m expecting fully
AI-generated movies by 2025.
Speech, Music, Protein, SWE, RPA
Generative AI is not limited to just text, image, and video—it
also includes speech, music, and protein-generation models. I
can't comment much on these areas since I haven't tried them out
myself. However, there are significant concerns about the
potential misuse of these technologies. For instance, scam calls
using cloned voices, face morphing for malicious purposes, and
generating fake videos of crimes someone never committed are
becoming real threats. Unfortunately, low-literacy individuals are
likely to be the most affected by these advancements, as they may
struggle to discern real content from AI-generated fakes.
Another important aspect to consider is the potential replacement
of software engineering jobs. So far, most software engineering
(SWE) tools haven't posed a significant threat to human
developers. AI tools like GitHub Copilot and Cursor are primarily
saving developers time rather than replacing them. The much-hyped
Devin, for instance, has been released with a hefty monthly cost
of $500 (₹42,000), making it inaccessible for many organizations.
This means junior developers are unlikely to be replaced anytime
soon, as the cost and limitations of these tools still make human
expertise indispensable.
Anthropic's computer agent and Microsoft's UI Parser might also
replace Robotic Process Automation tools like UiPath.
AI Politics
China has joined for the rise for AGI with the launch of Deepseek
R1. They are claiming they just used $5 million to train R1 while
OpenAI used $20million to train o1. Lack of GPUs lead China to
innovate in architectures.
While it is sceptical that the feud between Elon and Sam will slow
down the race in US. Elon is known for having beef with a wide
range of figures, including entreprenuers such as Zuck (Mark
Zuckerberg), Jeff (Bezos), Bill Gates, Yann LeCun, and left
politicians Geroge Soros, Biden, AOC, Nicolás Maduro (President of
Venezuela), Trudeau, and feminists, communists, and "woke"
activists. He even allegedly involved in 2024 US election by
manipulating X algorithm. While YouTube, Facebook and Instagram
were favoring the woke gang, X countered this using Community
Notes. Trump's election win has given Elon some leverage for his
companies.
Elon is triggered for not having a significant stake in OpenAI,
while Sam is accused of spreading "evil" by advocating for global
AI regulation. Elon is calling for OS even though what he done for
OS is releasing an unusably big Grok model. It's looks like Elon
doesn't want to live in a world where someone else making the
world a better place than himself.
Initially openAI open sourced GPT-2 and Whisper (speech to text
model). Now open source community is getting nothing from openAI.
Sam Altman also trying to the sole leader in the space by
promoting AI regulation around the world. European Union has fell
into that trap. Now things started to change after Trump's return.
Mistral's Le Chat's launch is a good comeback by Europe.
The open AI board battle which started at the end of 2023 led to
the collapse of the whole founding team. Ilya left to start SSI
(Safe Super Intelligence). Mira Murati left. So Sam Altman has
removed everyone who once defenestrated him from the company. Also
Andrej Karpathy who once was the lead of Tesla Self driving left
from there and joined OpenAI. He left from there also to start
youtube channel.
Nobel Prize
Nobel Prize for Physics and Chemistry was won by AI this year.
Geoffrey Hinton and John Hopefield for their contribution to
Neural Networks. Demis Hassabis (founder of Deepmind), John
Jumper and David Baker for their contributions to the
development of protein gen AI model.
AI in Indian film industry
AI usage in Indian movies are increasing. Vijay de-ageing in
GOAT, Vettaiyan movie Manassilaayo song used demised singer
voice. You can also see Unni Mukundan's voice in tamil and
telugu trailer of Marco is AI generated. Rekhachithram used AI
generated video of Mammootty.
OpenAI
OpenAI has released major things as a part of their 12 days of
Shipmas including o3 preview, sora, reinforcement finetuning,
o1, preference finetuning etc.
Google
Google has mogged OpenAI's shipmas event by announcing stuffs
such as Veo2 (video gen model), Gemini2 (Proprietory LLM),
PaliGemma2 (OS vision), Genie2 (foundational world model),
GenCast (weather predictor LLM), Alpha Qubit (Quandum Error
Correction Model), Willow (Quandum Chip).
Power Generation methods for AI
Sam Altman is invested in Helion Energy, a nuclear fusion power
generation startup. It's high time to small nuclear thermo tubes.
Major Things to look out for
- US vs China
- Can India build it's own soverign model?
-
Domain specific Indian models build by VSP (Coding for the
world, healthcare for India). VSP has got access to RTX 4000
NVIDIA GPU.
- AI generated movies
-
Dead internet theory (Social media flooded by AI generated
avatars, contents)
-
Powerful AI SWEs which can manipulate the whole repo in
VSCode/Cursor.