Shobhit Upadhyay

AI engineer in 365 days
▓▓▓▓░░░░░░░░░░░░░░ 124/365

[Shobhit Upadhyay @shoupxx]

It's been a month @Premium @Support @elonmusk I paid ₹4,700 for X Premium Annual via Google Play, but Premium was never activated. Google confirms my subscription is active, yet X support has only said it’s “raised internally” with no ETA
Weeks of support, no Premium, no refund.

[Shobhit Upadhyay @shoupxx]

Day 125 —> Becoming an AI Engineer

Implemented Instruction Fine-tuning with Alpaca-style data.

Instruction + Input → Output → Input/Target pairs (changes in Target pairs)

Insight: Fine-tuning aligns next-token prediction with task-specific behavior.

#LLM #FineTuning #GenAI

[Shobhit Upadhyay @shoupxx]

@sufiyan_404 Thanks man!!

[Shobhit Upadhyay @shoupxx]

Day 124 —> Becoming an AI Engineer

Built spam v/s Non spam :

Data → Preprocess → Batch → Load Weights → Classification Head → Fine-tune → Evaluate

Insight: Fine-tuning reuses learned representations while optimizing the model’s decision boundary for a new task.

#LLM #AI

[Shobhit Upadhyay @shoupxx]

@shivam74689 Great!! Proud of you buddy, not for just who you are, but how you are, always inspiring, always looking up for those dreams, which even everyone dares to imagine, maybe you failed to build, but you never failed to inspire everyone around you, this is just the beginning 🤝

[Shobhit Upadhyay @shoupxx]

@shivam74689 Good

[Shobhit Upadhyay @shoupxx]

Day 123 —> Becoming an AI Engineer

Building Spam vs Non-Spam classifier

Done with: Preprocessing → DataLoader → Pretrained Weights → Classification Head

Insight: A pretrained LLM can be repurposed for classification by adapting its output space to the target task

#LLM #DL

[Shobhit Upadhyay @shoupxx]

Day 122 —> Becoming an AI Engineer

LLM Fine-tuning:

Classification → specialized outputs
Instruction → follows task-specific prompts

Insight: Fine-tuning doesn’t rebuild the model—it specializes existing capabilities for a target behavior.

#LLM #FineTuning #GenAI

[Shobhit Upadhyay @shoupxx]

Day 121 —> Becoming an AI Engineer

Tried loading pre-trained GPT-2 (124M) weights locally.

~500 MB download at ~55 KiB/s… patience became part of the training loop

Insight: Working with LLMs isn’t just about arch. —data, weights, storage & compute are engineering problems too

[Shobhit Upadhyay @shoupxx]

@shivam74689 Good

[Shobhit Upadhyay @shoupxx]

Day 120 —> Becoming an AI Engineer

Temp. Scaling & Top-K Sampling for LLM generation.

Temperature → controls randomness
Top-K → limits candidate tokens

Insight: Decoding doesn’t change what the model knows—it controls how that knowledge is sampled into text

#LLM #GenAI #DL

[Shobhit Upadhyay @shoupxx]

@shivam74689 Good

[Shobhit Upadhyay @shoupxx]

@Officially_Dev_ Read the post carefully man!

[Shobhit Upadhyay @shoupxx]

Day 119 —> Becoming an AI Engineer

LLM from scratch, including --> full forward + loss + backprop pipeline.

GPT-2’s arch~164M parameters across embeddings, attention, FFNs & Transformer blocks.

Insight: Scale comes from simple operations × learned parameters.

#DL #LLM #GPT2

[Shobhit Upadhyay @shoupxx]

@SAkasxh Great

[Shobhit Upadhyay @shoupxx]

Day 118 —> Becoming an AI Engineer

My first small LLM . {Excluding Backprop. part}

Input → GPT → Generated Loss

Train Loss: 10.97
Val Loss: 10.98

Insight: Building an LLM makes the theory tangible — every token prediction is a learned probability distribution.

#LLM #GPT

[Shobhit Upadhyay @shoupxx]

@yur_o_0 Lmao

[Shobhit Upadhyay @shoupxx]

Day 117 —> Becoming an AI Engineer

Losses :

BCE → measures how well predicted probabilities match the target.
Perplexity → measures how uncertain the model is when predicting the next token.

Insight: Lower perplexity = better next-token predictions.

#LLM #GPT #DeepLearning

[Shobhit Upadhyay @shoupxx]

Day 116 —> Becoming an AI Engineer

How GPT generates new text:
Tokens → GPT → Logits → Softmax → Next Token → Repeat

Insight: GPT doesn’t generate a sentence at once—it builds it token by token, turning probability into text.

#LLM #GPT #GenAI

[Shobhit Upadhyay @shoupxx]

Day 115 —> Becoming an AI Engineer

Implemented the GPT architecture from scratch.

124.4M trainable params | ~622 MB

Shocked how much engineering sits behind a seemingly simple architecture.

Insights : Simple blocks + scale → powerful capabilities.

#LLM #GPT #DL #NLP