Ashish Patel 🇮🇳

Sr.Data Scientist & AI Consultant

Oracle

▪️ Led the AI strategy, architecture, and design for Artificial Intelligence, Machine Learning, and Data Science products, services, and solutions, aligning them with Oracle Cloud Applications, OCI, and Oracle Enterprise.
▪️ Provided thought leadership and technology oversight for AIML Development teams, directly influencing business and technology decisions.
▪️ Spearheaded the development and deployment of AI Agents for scalable, resilient enterprise applications, driving AI-powered automation across business functions.
▪️ Architected and implemented MCP server enterprise architecture, ensuring seamless integration with cloud-native environments and maximizing the scalability and performance of enterprise applications.
▪️ Collaborated with senior executives across multiple business lines to translate business requirements into technology solutions, aligning them with the overall AI strategy and next-gen tech stack.
▪️ Directed the development of machine learning models, integrating them with full-stack software ecosystems, and contributed to the strategic design of data engineering.
▪️ Worked with product managers and development teams to define project scopes and software requirements, ensuring the timely delivery of projects meeting quality standards.
▪️ Authored and maintained technical documentation, including architecture diagrams, design specifications, and system diagrams.
▪️ Led the adoption of cutting-edge technologies and frameworks (e.g., PyTorch, TensorFlow, Scikit-learn), driving innovation and best practices within the team.
▪️ Mentored junior developers, fostering skill development and encouraging a culture of creativity and ownership within the AIML domain.
▪️ Stayed at the forefront of industry trends, ensuring the team remained competitive by recommending new technologies and platforms.

Senior Principal Architect - AIML & Data Science

▪️Design and/or delivery of solutions, assets, and content at sustainable AI.
▪️Apply methodology and governance around project development.
▪️Document and articulate clear understanding of customer/partner success criteria.
▪️Apply critical thinking, technical knowledge, customer/partner insights to solve problems with industry methodologies, and frameworks.
▪️Collaborate with IBM and AWS specialists to design and build complex enterprise solutions with foundation of Red Hat Open Shift, AWS Infrastructure, and IBM Software technology.
▪️Standardize solutions in accordance with the AWS Well-Architected Framework across all pillars, including operational excellence, security, performance efficiency, cost optimization, and reliability.
▪️Define, lead, and execute architecture design workshops and white-boarding sessions to build implementation plans.
▪️Develop deep relationships with key customers through solution integration to drive long-term IBM and AWS adoption.
▪️Team well with our key IBM & Red Hat business units, and strategic partners to support our partnership.
▪️Consult as a subject matter expert on our IBM/AWS technology and services portfolio.
▪️Author use cases, reference architectures, blog posts, and other field and external content.
▪️Work in sandbox environments to demonstrate IBM Hybrid Cloud Capabilities with AWS.
▪️Led the design and automation of tasks in Oracle HCM using Generative AI and LLMs, streamlining HR processes and enhancing efficiency.
▪️Mentored junior and senior scientists, guiding their development and ensuring successful project execution.
▪️Developed production code and championed best coding and engineering practices in AI and machine learning projects.
▪️Identified and mitigated risks in AI project execution, ensuring successful business and technical outcomes.
▪️Collaborated with product managers and technical leaders to ensure timely and successful delivery of AI models and services.

Sr. AWS AI/ML Solution Architect

Cygnet Infotech

• Interact with customer and come up with use-case’s on ML/AI and use the same in
solution integration on existing business. 
• Selecting features, building, and optimizing classifiers using machine learning techniques.
• Develop and test Computer vision algorithms to edge devices and advanced hardware systems created.
• Develop and evaluate state-of-the-art computer vision algorithms for real-time edge devices
and other hardware devices.
• Perform technical performance benchmarking and analyses to support engineering decisions
• Data mining using state-of-the-art methods 
• Extending company’s data with third party sources of information when needed 
• Enhancing data collection procedures to include information that is relevant for building
analytic systems 
• Processing, cleansing, and verifying the integrity of data used for analysis.
• Participate in design reviews. 
• Doing ad-hoc analysis and presenting results in a clear manner 
• Creating automated anomaly detection systems and constant tracking of its performance 
• Research and devise innovative statistical models for data analysis 
• Work as the lead data strategist, identifying and integrating new datasets that can be 
leveraged through our product capabilities and work closely with the engineering team to
strategize and execute the development of data products. 
• Collaborate with product design and engineering to develop an understanding of needs
and Quality. 
• Analyse data for trends and patterns and Interpret data with a clear objective in mind. 
• Provide deep learning and machine learning base solutions to Solve varies Real world
problem.

Projects : Semantic Segmentation, Image classification,  3d Sementic Segmentation, Few-shot Image Classification, Object detection, Image Quality Assessment, Pose Estimation, Key-point Detection,  Image Augmentation, Super Resolution

Senior Data Scientist

SYMBIOSIS INTERNATIONAL UNIVERSITY

- Teaching Computer vision to M.tech Students.

Topics I teach: 

𝗕𝗮𝘀𝗶𝗰𝘀 𝗼𝗳 𝗗𝗶𝗴𝗶𝘁𝗮𝗹 𝗜𝗺𝗮𝗴𝗶𝗻𝗴
- Image Acquisition, Sampling, Quantization, Difference in Monochrome and Multichrome imaging, the concept of color spaces, point processing techniques, mask processing methods, image filtering, shape in images, edge detection, gradient operators- Roberts, Sobel, Prewitt, Canny, Slope magnitude method, morphological image processing, erosion, dilation, opening, closing, hit-n-miss transform, thinning, Top Hat transformation, Bottom hat transformation

𝗜𝗺𝗮𝗴𝗲 𝗥𝗲𝗽𝗿𝗲𝘀𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗥𝗲𝗴𝗶𝗼𝗻 𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀
 - Shape Descriptors-contour based, region-based, Boundary based; Thresholding based segmentation, Watershed based Segmentation, Gray level Co-occurrence Matrix-energy, entropy, maximum probability, contrast, correlation; wavelets, wavelet Pyramids, Image matching, 
- Hough Transform, RANSAC, Large-scale retrieval: Spatial Verification, TF-IDF, Query Expansion, feature encoding, Crowdsourcing, and Human Computation, Corners - Harris and Hessian Affine, Orientation Histogram, SIFT, SURF, HOG, GLOH, Scale-Space Analysis- Image Pyramids and Gaussian derivative filters, Gabor Filters and DWT.

Industry Associate Professor

Packt

TensorFlow Developer Certificate Guide : 

https://www.amazon.in/TensorFlow-Developer-Certificate-Comprehensive-Guide-ebook/dp/B0BVMQ1TZS

Technical Book Reviewer

Apress

Book Author

Manning Publications Co.

Prompt Softech

•	Collaborated with business teams to gather and analyze requirements, ensuring 100% alignment with organizational objectives and delivering solutions that met strategic goals.
•	Enhanced training datasets using advanced image augmentation techniques, improving model robustness and achieving a 15% increase in classification accuracy.
•	Leveraged Transfer Learning methodologies, repurposing pre-trained models to deliver 25% faster development cycles while achieving notable accuracy improvements across multiple tasks.
•	Implemented state-of-the-art deep learning models for image classification, extracting meaningful insights from complex datasets and enabling 98% accuracy in production systems.
•	Validated and fine-tuned models on extensive datasets containing thousands of images, ensuring 99% reliability and seamless deployment in real-world applications.

Softweb Solutions Inc.

•	Data mining using state-of-the-art methods. 
•	Hands-on experience in selecting features, building, and optimizing classifiers using machine-learning techniques. 
•	Enhancing data collection procedures to include information that is relevant for building analytic systems 
•	Processing, cleansing, and verifying the integrity of data used for analysis.
•	Doing ad-hoc analysis and clearly presenting results.
•	Creating anomaly detection systems and constant tracking of its performance.
•	Having strong knowledge in Statistics, Feature Engineering, Model Selection, Model Evaluation, and Feature Scaling to build accurate machine learning.
•	Experience in tuning and optimize ML/DL model for performance improvement.
•	Hands-on experience working with live projects on different machine learning techniques.
•	Having strong problem-solving capabilities.
•	Working with most demand industrial technology Framework like pandas, numpy, matplotlib, nltk(NLP), Keras(Deep Learning, tensor flow scikit learn(Machine learning), open cv

Projects: Image Classifications, Facial Expression Recognition, Recommendation system, Predictive Modeling, Regression Analysis

Solaris Publication

●	Research raw data and disseminate easily understood insights to provide organizational direction.
●	Produced a company guidelines and best practices document based on consumer insight analysis.
●	Performed regular web scraping to synthesize data.

Project: Image Classification, Seasonality Trends Analysis, Recommendation System

Jr. Data Scientist

🔥 6x Linkedln Top Voice | AI Research Scientist & Chief Data Scientist at IBM | Generative AI Expert | Author - Hands-on Time Series Analytics with Python | IBM Quantum ML Certified | 11+ Years in AI | MLOps | IIMA |

I have over 12+ years, Author, Data Scientist and Researcher with 8+ Years of Experience of Data Science technology and Research experience in wide functions including predictive modelling, data preprocessing, feature engineering, machine learning and deep learning. Currently, I work as Sr.Aws AI ML Solution Architect(Chief Data Scientist) at IBM India Pvt Ltd. Helps to Collaborate with IBM and AWS specialists to design and build complex enterprise solutions with the foundation of Red Hat Open Shift, AWS Infrastructure, and IBM Software technology. Standardize solutions in accordance with the AWS Well-Architected Framework across all pillars, including operational excellence, security, performance efficiency, cost optimisation, and reliability.

If you catch me outside of work, I'm an adventurous traveller and avid cricket lover!

𝗖𝗼𝗺𝗽𝗲𝘁𝗲𝗻𝗰𝗶𝗲𝘀: Data science, Machine learning, Deep learning, Audio Processing, Python, Computer Vision, SQL, Keras, TensorFlow, Librosa, Pyaudio, Mlflow, tube flow, mlops, DVC, MLOps on AWS, Kubeflow on AWS, 

Enterprise LLMs Services: Amazon Bedrock, Amazon Titan, IBM WatsonX.ai, Azure OpenAI, NvidiaNemo, Google PaLM

LLMs Knowledge: GPT-4, Claude, ChatGPT, OpenChat, UltraLM, OpenCoderPlus, WizardLM, Nous Hermes, Vicuna, Baize-v2, Guanaco, Minotaur, Davinci003, Falcon, Alpaca Farm PPO Sim (GPT-4), Pythia, Cohere, Alpaca Farm PPO Human, Cohere Chat, Alpaca, Falcon 7B Instruct, Davinci001.

MLOps on AWS Services: Amazon SageMaker, AWS CodePipeline, AWS CodeBuild, AWS CodeCommit, AWS CodeDeploy, AWS Batch, AWS Glue, AWS Data Pipeline, Amazon Elastic Container Service, AWS Farget, AWS Lambda, AWS EKS(Elastic Kubernetes Service), AWS ECS(Elastic Container Service), AWS APP Runner, AWS Cloud Formation, AWS ECR(Elastic Container Registry), AWS CloudWatch, AWS Auto Scaling, AWS CloudTrail, AWS EC2, AWS EBS(Elastic Block Store), AWS S3, AWS Route53

I also like meeting new people over a cup of coffee. Feel free to reach out at ashishpatel.ce.2011@gmail.com 

I am a passionate blogger, Researcher like to write about Machine learning and deep learning related Articles. you can visit my blog 

𝗟𝗶𝗻𝗸: https://medium.com/ml-research-lab.
book: https://www.amazon.in/B-V-Vishwas-ebook/dp/B08GLG46PQ/
Orcid Profile: https://orcid.org/0000-0002-5385-3897
Google Scholar : https://scholar.google.com/citations?hl=en&authuser=1&user=MfHiCbkAAAAJ

P.S. If you're passionate about creating tech-focused products to help people reach their potential, I am happy to connect at https://www.linkedin.com/in/ashishpatel2604

Sr. AWS AI/ML Solution Architect(Chief Data Scientist)

▪️Design and/or delivery of solutions, assets, and content at sustainable AI
▪️Apply methodology and governance around project development.
▪️Document and articulate clear understanding of customer/partner success criteria.
▪️Apply critical thinking, technical knowledge, customer/partner insights to solve problems with industry methodologies, and frameworks.
▪️Collaborate with IBM and AWS specialists to design and build complex enterprise solutions with foundation of Red Hat Open Shift, AWS Infrastructure, and IBM Software technology.
▪️Standardize solutions in accordance with the AWS Well-Architected Framework across all pillars, including operational excellence, security, performance efficiency, cost optimization, and reliability.
▪️Define, lead, and execute architecture design workshops and white-boarding sessions to build implementation plans.
▪️Develop deep relationships with key customers through solution integration to drive long-term IBM and AWS adoption.
▪️Team well with our key IBM & Red Hat business units, and strategic partners to support our partnership.
▪️Consult as a subject matter expert on our IBM/AWS technology and services portfolio.
▪️Author use cases, reference architectures, blog posts, and other field and external content.
▪️Work in sandbox environments to demonstrate IBM Hybrid Cloud Capabilities with AWS.

If you want to truly understand LLMs…
try calculating them by hand first.

That’s the magic of My friend Prof. Tom Yeh  Deep Learning Math Workbook.

It forces you to slow down.
Do the math.
See the mechanics.
Understand what’s happening behind every token prediction.

Here’s what each chapter teaches you — and why it matters for LLMs:

1. Dot Product

When you calculate it manually, you see how “similarity” is measured.
In LLMs → this becomes attention scores.

2. Matrix Multiplication

Doing it by hand shows how information combines and transforms.
In LLMs → this powers every forward pass through every layer.

3. Linear Layer

Manually computing W·X + b shows how inputs turn into features.
In LLMs → this is how Q/K/V projections and MLPs work.

4. Activation

Calculating ReLU, Sigmoid, Tanh makes you feel the effect of non-linearity.
In LLMs → this gives the model its reasoning ability (GELU, gates, etc.).

5. Artificial Neuron

Working through a single neuron reveals the logic inside every unit.
In LLMs → millions of these form the model’s internal understanding.

6. Batch

Manually applying operations across multiple examples shows training efficiency.
In LLMs → batching is the secret to stable, scalable learning.

7. Connection

Seeing residuals and wiring by hand makes network flow click.
In LLMs → this prevents memory loss and exploding gradients.

8. Hidden Layer

When you compute hidden transformations, you see abstraction forming.
In LLMs → this is where meaning and representation live.

9. Deep

Stacking operations manually shows how depth compounds understanding.
In LLMs → deeper stacks = richer reasoning and context retention.

10. Wide

Calculating wide layers reveals how models store more information.
In LLMs → width boosts knowledge capacity and expressiveness.

11. Softmax

Manually normalizing logits shows how models “choose” the next token.
In LLMs → this becomes the probability distribution we decode from.

12. Gradient

Computing gradients by hand shows how learning actually happens.
In LLMs → gradients tune billions of weights during training.

When you calculate these steps manually, you stop memorizing formulas…
and start understanding the machine.

That’s the real gift of Prof. Tom Yeh’s workbook — it makes deep learning visible, tangible, and uncomplicated.

♻️ Repost this if you want more people to build real intuition — not just run code.

P.S. Which chapter should I break down next with a simple example?

No alternative text description for this image

Everyone is yelling about “AGI” right now, but the people closest to the frontier are still giving wide, uncertain timelines.

Ilya just said 5–10 years on the Dwarkesh Patel podcast.
Demis Hassabis has said similar numbers.
Meanwhile, expert surveys still place the median expectation closer to the 2040s–2050s, and a good chunk of researchers think it could be much later — or never.

there is no consensus, just speculation.

That’s why this meme fits the current moment: companies barking at each other about AGI while no one has crossed the threshold.

The real progress isn’t in “AGI claims.”
It’s in the unglamorous fundamentals:

• reliability under real workloads
• reasoning that actually generalizes
• efficient multimodal integration
• scalable data + compute strategies
• products that deliver value today, not in hypothetical AGI futures

The noise is about AGI timelines.
The work is about making current AI systems genuinely useful.

🌡️ "LLM Temperature” is NOT the little slider you think it is.

Ever adjusted the temperature in an LLM? 
It’s a powerful setting that shapes LLM Model behavior, but it's often misunderstood. 

Let’s break down six key control of temperature and how they can impact model performance. 🔥

🔎 𝗪𝗵𝗮𝘁 𝗜𝘀 𝗧𝗲𝗺𝗽𝗲𝗿𝗮𝘁𝘂𝗿𝗲 𝗶𝗻 𝗟𝗟𝗠𝘀?

Temperature adjusts the probability distribution during inference:

 • 𝗛𝗶𝗴𝗵𝗲𝗿 𝘁𝗲𝗺𝗽𝗲𝗿𝗮𝘁𝘂𝗿𝗲 = More random, creative output.
 • 𝗟𝗼𝘄𝗲𝗿 𝘁𝗲𝗺𝗽𝗲𝗿𝗮𝘁𝘂𝗿𝗲 = More predictable, factual output.

But here’s the thing—temperature isn't just a simple dial. It plays six key roles in LLMs. Here’s how each one works:

🛠️ 𝗧𝗵𝗲 𝟲 𝗧𝘆𝗽𝗲𝘀 𝗼𝗳 𝗧𝗲𝗺𝗽𝗲𝗿𝗮𝘁𝘂𝗿𝗲 𝗖𝗼𝗻𝘁𝗿𝗼𝗹

1️⃣ 𝗦𝗮𝗺𝗽𝗹𝗶𝗻𝗴- 𝗧
Purpose: Controls randomness.
Use-case: Balances creativity vs. accuracy (e.g., chatbots vs. technical writing).
Tip: Increase for creative output, decrease for factual accuracy.

2️⃣ 𝗗𝗶𝘀𝘁𝗶𝗹𝗹𝗮𝘁𝗶𝗼𝗻-𝗧
Purpose: Helps smaller models learn from larger ones.
Use-case: Efficient model compression.

3️⃣ 𝗖𝗮𝗹𝗶𝗯𝗿𝗮𝘁𝗶𝗼𝗻-𝗧
Purpose: Adjusts model confidence.
Use-case: Important in regulated industries (finance, healthcare).

4️⃣ 𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻-𝗧
Purpose: Fine-tunes self-attention in transformers.
Use-case: Stabilizes complex sequences.

5️⃣ 𝗗𝘆𝗻𝗮𝗺𝗶𝗰 / 𝗔𝗱𝗮𝗽𝘁𝗶𝘃𝗲 -𝗧
Purpose: Adjusts temperature during token generation for a mix of creativity and accuracy.
Use-case: Perfect for interactive models like chatbots.

6️⃣ 𝗟𝗮𝘆𝗲𝗿-𝗪𝗶𝘀𝗲-𝗧
Purpose: Applies different temperatures across layers of the model.
Use-case: Preserves syntax in early layers, enhances semantics in later layers.

📆 𝗧𝗶𝗺𝗲𝗹𝗶𝗻𝗲 𝗼𝗳 𝗧𝗲𝗺𝗽𝗲𝗿𝗮𝘁𝘂𝗿𝗲 𝗶𝗻 𝗔𝗜

 • 1950s: Temperature concept originates in physics.
 • 1980s: Geoff Hinton adapts it for neural networks.
 • 2015: Temperature scaling introduced in knowledge distillation.
 • 2019-2023: Sampling temperature becomes standard in generative models (GPT-2, GPT-3).
 • 2023+: Six distinct types of temperature emerge in training and inference pipelines.

📊 𝗪𝗵𝘆 𝗧𝗲𝗺𝗽𝗲𝗿𝗮𝘁𝘂𝗿𝗲 𝗖𝗼𝗻𝘁𝗿𝗼𝗹 𝗠𝗮𝘁𝘁𝗲𝗿𝘀

𝗗𝗶𝗱 𝘆𝗼𝘂 𝗸𝗻𝗼𝘄? 

 • Even a 𝟬.𝟭-𝗱𝗲𝗴𝗿𝗲𝗲 shift in sampling temperature can cause a 𝟬.𝟴% 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆 𝗹𝗼𝘀𝘀 𝗳𝗼𝗿 𝗲𝘃𝗲𝗿𝘆 𝟭,𝟬𝟬𝟬 𝘁𝗼𝗸𝗲𝗻𝘀 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲𝗱! 
 • Temperature adjustments can dramatically change model output, from creativity to coherence.

💬 Let’s Discuss: Have You Experimented with Temperature?

Which temperature control has made the most difference in your models? Drop your thoughts below! 👇

I have over 10.5+ years, Author, Data Scientist and Researcher with 6+ Years of Experience of Data Science technology and Research experience in wide functions including predictive modelling, data preprocessing, feature engineering, machine learning and deep learning. Currently, I work as Sr.Aws AI ML Solution Architect(Chief Data Scientist) at IBM India Pvt Ltd. Helps to Collaborate IBM and AWS specialists to design and build complex enterprise solutions with foundation of Red Hat Open Shift, AWS Infrastructure, and IBM Software technology. Standardize solutions in accordance with the AWS Well-Architected Framework across all pillars, including operational excellence, security, performance efficiency, cost optimisation, and reliability.

Previously, I worked at Cygnet infotech Pvt Ltd as a Data Scientist, Improving products and services for our customers by using advanced analytics, standing up big-data analytical tools, creating and maintaining models, and onboarding compelling new data sets.

If you catch me outside of work, I'm an adventurous traveler and an avid lover of cricket!

𝗖𝗼𝗺𝗽𝗲𝘁𝗲𝗻𝗰𝗶𝗲𝘀: Data science, Machine learning, Deep learning, Audio Processing, Python, Computer Vision, SQL, Keras, TensorFlow, Librosa, Pyaudio, Mlflow, kubeflow, mlops, DVC, MLOps on AWS, Kubeflow on AWS.

MLOps on AWS Services: Amazon SageMaker, AWS CodePipeline, AWS CodeBuild, AWS CodeCommit, AWS CodeDeploy,AWS Batch, AWS Glue, AWS Data Pipeline, Amazon Elastic Container Service, AWS Farget, AWS Lambda, AWS EKS(Elastic Kubernets Service), AWS ECS(Elastic Container Service), AWS APP Runner, AWS Cloud Formation, AWS ECR(Elastic Container Registry), AWS CloudWatch, AWS Auto Scaling, AWS CloudTrail, AWS EC2, AWS EBS(Elastic Block Store), AWS S3, AWS Route53

I also like meeting new people over a cup of coffee. Feel free to reach out at ashishpatel.ce.2011@gmail.com 

I am a passionate blogger, Researcher like to write about Machine learning and deep learning related Articles. you can visit my blog 

𝗟𝗶𝗻𝗸: https://medium.com/ml-research-lab.
book: https://www.amazon.in/B-V-Vishwas-ebook/dp/B08GLG46PQ/
Orcid Profile: https://orcid.org/0000-0002-5385-3897
Google Scholar : https://scholar.google.com/citations?hl=en&authuser=1&user=MfHiCbkAAAAJ

P.S. If you're passionate about creating tech-focused products to help people reach their potential, I am happy to connect at https://www.linkedin.com/in/ashishpatel2604

Last week didn’t just feel big.I had the honour of inaugurating the upGrad Ahmedabad Learning Centre…

…and as an Amdavadi, this moment hit differently.
(It felt like the city is finally stepping into the future.)

But the real spark?

The conversations with the students.

The excitement.
The curiosity.
The obsession with AI.

It was absolutely electric.

And the questions they asked?

Not basic. Not surface-level.
But sharp, forward-thinking, “this generation is built different” kind of questions:

• “How do we build a career in AI if we’re starting from scratch?”
• “What skills in AI will matter the most in the next 3–5 years?”
• “Will AI replace jobs — or create even bigger opportunities?”
• “How do we stay relevant when AI tools evolve so fast?”
• “What does the industry actually expect from an AI fresher?”

These weren’t just questions.
They were signals.

Signals that Gujarat isn’t just catching up —
Gujarat is ready to lead.

And that’s exactly why this centre matters so much.

Here’s what made the day unforgettable:

1️⃣ Meeting Shubham Singh  & Ravneet Arora  
Two leaders who are fearlessly pushing to bring future-ready learning to Gujarat.

2️⃣ Feeling the energy in the room with 
Students. Working professionals. 
Tech enthusiasts.
Everyone driven by one thing — growth.

3️⃣ What this centre unlocks
• Data Science & AI/ML
• Full Stack Development
• Data Analytics
• Digital Marketing

These aren’t just courses.
They’re career accelerators.
Gateways to opportunities that didn’t exist a decade ago.

#Gujarat’s learners are ambitious.
They’re curious.
They’re hungry.

All they needed was a launchpad.

And now — they have one.

If you’re in #Ahmedabad and want to build a future-proof tech or AI career, this is your moment to step in.

♻️ Reshare this to inspire someone who needs to see it.

🚀 I finally got Kimi-K2-Thinking running locally.

And yes — it actually works.

Here’s what I learned (and why I’m impressed) 👇

1️⃣ It’s ridiculously smart
Kimi-K2-Thinking nails reasoning, coding, and agentic tasks.
Feels like GPT-4-tier thinking — but fully open.

2️⃣ I’m running it on my own hardware
No API latency. No privacy trade-offs.
The Unsloth Dynamic 1.8-bit quant slashed model size from 1TB → 230GB.
That’s an 80% reduction without killing accuracy.

3️⃣ Setup was easier than expected
Just swapped in the new GGUFs:
unsloth/Kimi-K2-Thinking-GGUF
and ran it through llama.cpp with:
--special → brings back the <think> tags.

4️⃣ Specs that actually work
I’m getting ~1–2 tokens/s on a single 24GB GPU.
If you’ve got ~250GB combined RAM+VRAM, it flies (5+ tokens/s).
Even with less, disk offloading keeps it functional.

5️⃣ We even fixed the tokenizer bug
Worked with the Kimi team to patch the missing system prompt.
Now it behaves like GPT-4o — clean, stable, consistent.

If you’re into local LLMs, this one’s worth your time.
Feels like the first real “thinking” model you can actually own.

♻ Repost this for your AI dev network.
Let’s push open-source forward.

P.S. I found this all configs + downloads here:
https://lnkd.in/dGUTS3wf

KV-Cache Drift: The Silent “Memory Smudge” Making Your LLM Dumber Every Night 🧠➡️🤏

Ever notice your LLM starts strong, but after a few interactions, it drifts off-topic? 
That’s 𝗞𝗩-𝗰𝗮𝗰𝗵𝗲 𝗱𝗿𝗶𝗳𝘁 in action. 

Here’s what’s happening:

𝗪𝗵𝗮𝘁 𝗶𝘀 𝗞𝗩-𝗖𝗮𝗰𝗵𝗲 𝗗𝗿𝗶𝗳𝘁?

Transformer models store 𝗞𝗲𝘆-𝗩𝗮𝗹𝘂𝗲 𝘃𝗲𝗰𝘁𝗼𝗿𝘀 to avoid recalculating every token. But, over time, when these “scratchpads” sit in 𝗚𝗣𝗨 𝗥𝗔𝗠, rounding errors and new attention patterns slowly 𝗿𝗲𝘄𝗿𝗶𝘁𝗲 𝘁𝗵𝗼𝘀𝗲 𝘃𝗲𝗰𝘁𝗼𝗿𝘀.

The result?
• 𝗙𝗶𝗿𝘀𝘁 𝗿𝗲𝗽𝗹𝘆: Perfect.
• 500𝘁𝗵 𝗿𝗲𝗽𝗹𝘆: Not so much.
This isn’t a bug. It's a 𝗳𝘂𝗻𝗱𝗮𝗺𝗲𝗻𝘁𝗮𝗹 𝗺𝗲𝗺𝗼𝗿𝘆 𝗶𝘀𝘀𝘂𝗲 in LLMs.

𝗪𝗵𝘆 𝗦𝗵𝗼𝘂𝗹𝗱 𝗬𝗼𝘂 𝗖𝗮𝗿𝗲? 🤔

If you’re working with:
• 𝗦𝘁𝗿𝗲𝗮𝗺𝗶𝗻𝗴 𝗔𝗣𝗜𝘀 that hold sessions for days
• 𝗟𝗼𝗻𝗴-𝗿𝘂𝗻𝗻𝗶𝗻𝗴 𝗰𝗼𝗱𝗲 𝗮𝗴𝗲𝗻𝘁𝘀 across CI pipelines
• 𝗟𝗮𝗿𝗴𝗲-𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗥𝗔𝗚 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 (100k+ tokens)
• 
Then 𝗞𝗩-𝗰𝗮𝗰𝗵𝗲 𝗱𝗿𝗶𝗳𝘁 is quietly affecting your model’s 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆, 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮𝗻𝘆 𝗽𝗮𝗿𝗮𝗺𝗲𝘁𝗲𝗿 𝘂𝗽𝗱𝗮𝘁𝗲𝘀.

𝗔 𝗗𝗲𝗲𝗽𝗲𝗿 𝗟𝗼𝗼𝗸 𝗮𝘁 𝘁𝗵𝗲 𝗣𝗿𝗼𝗯𝗹𝗲𝗺 👀

KV-cache drift isn’t new. 

👉  1980s: RNNs suffered “vanishing memory.”
👉  2017: “Attention Is All You Need” kills recurrence, but introduces KV vectors.
👉  2019: KV-cache becomes default optimisation (50 % latency cut).
👉  2023: 100k-context models → caches live for days.
👉  2024: Papers (Elastic-Cache, GEAR) measure drift: deeper layers shift up to 0.8 % per 1k tokens.
👉  2025: Fixes land — periodic refresh, rotary-corrected positions, 8-bit KV quant with error-feedback.

𝗪𝗵𝗮𝘁 𝗖𝗮𝗻 𝗬𝗼𝘂 𝗗𝗼 𝗔𝗯𝗼𝘂𝘁 𝗜𝘁 𝗧𝗼𝗱𝗮𝘆? 🚀

Good news: fixes are on the way.
• 𝗘𝗹𝗮𝘀𝘁𝗶𝗰-𝗖𝗮𝗰𝗵𝗲: A refresh strategy that updates only the parts that need it. 𝗦𝗽𝗲𝗲𝗱 𝘂𝗽 your model 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗿𝗲𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴.
• 𝗚𝗘𝗔𝗥: 4-bit quantization + outlier correction for 𝗺𝗮𝘀𝘀𝗶𝘃𝗲 𝗺𝗲𝗺𝗼𝗿𝘆 𝗮𝗻𝗱 𝘀𝗽𝗲𝗲𝗱 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁𝘀, with near-lossless accuracy.
• 𝗥𝗮𝘇𝗼𝗿𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻: Cuts cache size by 70% for certain tokens, reducing unnecessary drift.

𝗬𝗼𝘂𝗿 𝗔𝗰𝘁𝗶𝗼𝗻𝗮𝗯𝗹𝗲 𝗣𝗹𝗮𝗻 📝

1. 𝗧𝘂𝗿𝗻 𝗼𝗻 𝗘𝗹𝗮𝘀𝘁𝗶𝗰-𝗖𝗮𝗰𝗵𝗲: Just a quick patch.
2. 𝗔𝗱𝗱 𝗚𝗘𝗔𝗥 4-𝗯𝗶𝘁 𝗰𝗼𝗺𝗽𝗿𝗲𝘀𝘀𝗶𝗼𝗻: Perfect for memory-heavy pods.
3. 𝗖𝗼𝗺𝗯𝗶𝗻𝗲 𝘄𝗶𝘁𝗵 𝗥𝗮𝘇𝗼𝗿𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻: If your model serves >70𝗸 𝘁𝗼𝗸𝗲𝗻𝘀.
4. 𝗦𝗰𝗵𝗲𝗱𝘂𝗹𝗲 𝗳𝘂𝗹𝗹-𝗰𝗮𝗰𝗵𝗲 𝗿𝗲𝗳𝗿𝗲𝘀𝗵𝗲𝘀: During off-peak hours or after processing N tokens.
The result? No more KV-cache drift. Your model stays sharp, 𝗻𝗼 𝗺𝗮𝘁𝘁𝗲𝗿 𝗵𝗼𝘄 𝗹𝗼𝗻𝗴 𝗶𝘁’𝘀 𝗿𝘂𝗻𝗻𝗶𝗻𝗴. ✨

𝗣.𝗦. Have you encountered KV-cache drift in your LLMs? How did you tackle it? Drop your thoughts below! 👇

Shekar Ramachandran

🕰️ Temporal Reasoning: LLMs struggle with time, sequence, and change.

𝗪𝗵𝗮𝘁 𝗶𝗳 𝗜 𝘁𝗼𝗹𝗱 𝘆𝗼𝘂 𝘁𝗵𝗮𝘁 𝗔𝗜 𝗰𝗮𝗻 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲 𝘁𝗲𝘅𝘁, 𝘄𝗿𝗶𝘁𝗲 𝗰𝗼𝗱𝗲, 𝗮𝗻𝗱 𝗲𝘃𝗲𝗻 𝗮𝗻𝘀𝘄𝗲𝗿 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻𝘀, 𝗯𝘂𝘁 𝗶𝘁 𝘀𝘁𝗶𝗹𝗹 𝘀𝘁𝗿𝘂𝗴𝗴𝗹𝗲𝘀 𝘄𝗶𝘁𝗵 𝘀𝗼𝗺𝗲𝘁𝗵𝗶𝗻𝗴 𝘀𝗶𝗺𝗽𝗹𝗲: 

𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 𝘁𝗶𝗺𝗲? When it comes to knowing 𝘸𝘩𝘦𝘯 something happened, or in what order, AI gets lost.

While GPT-4 and other LLMs are great at recalling facts, 𝘁𝗲𝗺𝗽𝗼𝗿𝗮𝗹 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 — understanding the timing and sequence of events — is still a massive challenge.

⏳ 𝗪𝗵𝗮𝘁 𝗶𝘀 𝗧𝗲𝗺𝗽𝗼𝗿𝗮𝗹 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴?

𝗦𝗶𝗺𝗽𝗹𝘆 𝗽𝘂𝘁: Temporal reasoning is the ability to understand 𝘸𝘩𝘦𝘯 things happen and how they relate over time.

For example:
• “The iPhone 12 was released in 2020.”
• “The iPhone 15 came after it.”
• “The iPhone 12 is the older model.”
This is second nature for humans, but for AI? It's a struggle.

📚 𝗔 𝗕𝗿𝗶𝗲𝗳 𝗛𝗶𝘀𝘁𝗼𝗿𝘆 𝗼𝗳 𝗧𝗲𝗺𝗽𝗼𝗿𝗮𝗹 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗖𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲𝘀 𝗶𝗻 𝗟𝗟𝗠𝘀
• 1980𝘀: Temporal reasoning frameworks like Allen’s interval algebra laid the foundation.
• 2020𝘀: Transformer models like GPT-3 and GPT-4 revealed AI's temporal limitations.
• 2021: Datasets like 𝗧𝗶𝗺𝗲𝗤𝗔 tested AI's ability to handle time-sensitive data.
• 2023–𝗣𝗿𝗲𝘀𝗲𝗻𝘁: Benchmarks like 𝗧𝗲𝗺𝗽𝗥𝗲𝗮𝘀𝗼𝗻 and 𝗖𝗵𝗿𝗼𝗻𝗼𝗦𝗲𝗻𝘀𝗲 are helping improve AI’s temporal understanding.

🧠 𝗪𝗵𝘆 𝗶𝘀 𝗧𝗲𝗺𝗽𝗼𝗿𝗮𝗹 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗛𝗮𝗿𝗱 𝗳𝗼𝗿 𝗟𝗟𝗠𝘀?
• 𝗙𝗿𝗼𝘇𝗲𝗻 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲: LLMs know facts until their last training update, but can't track when facts become outdated.
• 𝗡𝗼 𝗜𝗻𝘁𝗲𝗿𝗻𝗮𝗹 𝗖𝗹𝗼𝗰𝗸: AI lacks a built-in sense of time.
• 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗚𝗮𝗽𝘀: Models lose track of events if they’re outside their processing window.

🚀 𝗛𝗼𝘄 𝘁𝗼 𝗙𝗶𝘅 𝗜𝘁
1. 𝗧𝗲𝗺𝗽𝗼𝗿𝗮𝗹 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗚𝗿𝗮𝗽𝗵𝘀 (𝗧𝗞𝗚𝘀): Connect facts to specific times.
2. 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹-𝗔𝘂𝗴𝗺𝗲𝗻𝘁𝗲𝗱 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 (𝗥𝗔𝗚): Use real-time data to provide up-to-date answers.
3. 𝗖𝗼𝗻𝘁𝗶𝗻𝘂𝗮𝗹 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴: Allow models to update their knowledge over time.

📌 𝗧𝗵𝗲 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆

AI is incredible at recalling facts, but 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 𝘁𝗵𝗲 𝘁𝗶𝗺𝗶𝗻𝗴 𝗼𝗳 𝘁𝗵𝗼𝘀𝗲 𝗳𝗮𝗰𝘁𝘀 is still a major challenge. 
The future of AI isn’t just about knowing 𝘸𝘩𝘢𝘵 happened; it's about knowing 𝘄𝗵𝗲𝗻 it happened.

𝗣.𝗦. 𝗛𝗼𝘄 𝗱𝗼 𝘆𝗼𝘂 𝘁𝗵𝗶𝗻𝗸 𝗔𝗜 𝗰𝗮𝗻 𝗶𝗺𝗽𝗿𝗼𝘃𝗲 𝗶𝘁𝘀 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 𝗼𝗳 𝘁𝗶𝗺𝗲?

Research Paper collection in comments 👇.
Drop your thoughts below 👇 and reshare if you found this insight helpful! 🔁

I learned something wild listening to an O’Reilly podcast: we’re only 5% of the way to real LLMOps maturity.

The guest was Abi Aryan ☯︎  of LLMOps.
Here are the biggest lessons I walked away with👇

⸻

1️⃣ LLMOps isn’t MLOps 2.0

It’s a different world.

• From deterministic → nondeterministic
• From “deploy the model” → manage distributed inference
• From accuracy → latency + throughput + memory + cost

And honestly?
Software engineers seem to transition faster than MLOps engineers —
because LLMOps feels more like distributed systems than classic ML.

⸻

2️⃣ Agent systems = distributed systems (not magic)

Modern setups usually have:

• A planner
• Specialized workers
• A reassembly layer

But this creates real problems:

• GPU load balancing
• Consensus when an agent fails
• Errors multiplying across the chain

Also… most “agents” today are just fancy if-this-then-that workflows.

⸻

3️⃣ Observability needs a new mindset

It’s not about watching dashboards.
It’s about anomaly detection.

• What changed across 50+ signals?
• What’s unusual before an error appears?
• What’s drifting?

It’s more like monitoring fraud detection than monitoring ML.

⸻

4️⃣ FinOps is the silent pain point

We’re maybe 5% of the way to good tooling.

Real cost buckets:

1️⃣ Data engineering
2️⃣ Compute-bound work
3️⃣ Communication-bound bottlenecks (GPU↔GPU)

Most teams overspend because they:

• Use reasoning models unnecessarily
• Overfill KV cache
• Optimize compute when the real issue is memory

⸻

5️⃣ RL has a fatal flaw

LLM RL environments are static…
but agents need dynamic learning.

Once a model “finishes” the environment — it plateaus.
It’s the same freshness problem we see in RAG.

Curriculum learning exists in research,
but no practical implementation yet.

Huge gap.

⸻

6️⃣ The future (and how I’m thinking about my own skills)

Areas that feel massively underserved:

• System-level profiling
• Dynamic RL environments
• Data engineering (big practical wins)

And the skill order that makes sense:

• Learn distributed systems
• Learn profiling (memory, compute, comms)
• Then go deeper into model internals

LLMs aren’t “another ML model.”
They’re Software 3.0 — end-to-end systems where the model is the product.

If this breakdown helped, consider resharing ♻️

P.S. What part of LLMOps do you think is most underserved right now?

Google just challenged backprop.
It’s called Nested Learning — and it quietly changes how AI actually learns. 🧠

🔍 The hidden flaw in today’s LLMs

Every model — GPT-4, Claude, Gemini — runs on one global optimization loop.

One loss.
One gradient flow.
Everything updates together.

That’s why fine-tuning feels dangerous —
new gradients overwrite old memories.

We’ve built intelligence that learns fast ⚡️
but forgets even faster.

⚙️ What Google did differently

Nested Learning breaks that monolith apart.

Each layer (even each attention head) gets its own local optimizer —
its own context,
its own update rate,
its own objective.

Local loops learn fast 🌀
Outer loops consolidate slow 🧩

The network becomes a hierarchy of cooperating learners,
not one centralized brain.

It’s deep learning → re-engineered for memory.

🚀 Why it matters

1️⃣ Continual learning — new data doesn’t erase old skills.
2️⃣ Memory efficiency — no replay buffers or retrains.
3️⃣ Long-context reasoning — inner loops adapt locally, outer loops stabilize globally.
4️⃣ Biological parallel — fast plasticity, slow consolidation.

It’s not scaling.
It’s evolution.

💭 My take

If Nested Learning works at scale,
we move from static models that store knowledge
to dynamic systems that preserve and evolve it.

That’s not more compute.
That’s architectural intelligence.

♻️ Repost if you think the next AI revolution won’t come from bigger models —
but from smarter learning itself.

P.S. Would you bet on backprop still ruling in 2030? 🤔

🔍 Top-P Bleed: When Your LLM Starts Forgetting Things 🩸

Ever seen your LLM stop mid-sentence?
Forget how to close a bracket?
Or just… trail off?

That’s not a glitch.
It’s Top-P Bleed.

🩸 What’s Going On

When your model generates text, it picks the next word based on probability —
only from the “most likely” ones (say the top 95%).

But as the sentence grows, those probabilities keep shifting.
And words that were “barely included” early on can slip out of range later.

The result?
Important tokens disappear.
Your model “bleeds” context.

(Think: an open { that never finds its }.)


⚡ Why You Should Care

Because that tiny drift shows up everywhere:

•	In code: missing braces or parentheses → broken logic.
•	In SQL: dropped WHERE or quotes → data chaos.
•	In legal text: lost “not” or “except” → reversed meaning.
•	In docs: half-written formulas → credibility crash.

Small probability errors.
Big real-world impact.


🕰️ How We Got Here

•	2019: Researchers introduced Top-P sampling — smoother than older methods.
•	2021–22: It became standard in models like GPT-3.
•	2023–24: Studies uncovered “entropy buildup” — a slow drift that warps token choices over time.

That’s when semantic decay began sneaking into long outputs.


🛠️ Smart Fixes

1.	Dynamic Top-P Refresh
Re-evaluate the nucleus after every word.
Keeps key tokens (like punctuation) in play.
2.	Min-P Safety Net
Set a minimum threshold so rare but vital words don’t vanish.
3.	Hybrid Top-K + Top-P
Use both filters together for better structure.
(Yes — GitHub Copilot already does this.)
4.	Entropy-Aware Sampling
When uncertainty rises, expand the options.
Translation → let the model breathe instead of choke.

💡 The Big Idea

Top-P Bleed isn’t a bug.
It’s probability drift in motion.

Your model’s “nucleus” isn’t fixed — it moves and reshapes with every word.

If you want cleaner, longer, more reliable outputs:
✅ Use dynamic or hybrid top-p methods
✅ Log per-token probabilities
✅ Adjust temperature to stabilize creativity

💬 Ever had your LLM “forget” to finish a thought?
Share your debugging horror stories below 👇

♻ Repost this if you’ve ever watched your AI trail off mid-sentence and thought, “Wait… what just happened?”

P.S. It wasn’t hallucination. It was probability bleed.

Catastrophic Forgetting’s 𝗦𝗰𝗮𝗹𝗲 𝗟𝗮𝘄: Why Bigger LLMs Forget Faster 🚨

Ever fine-tuned a model and watched it forget everything? 🤯 The bigger the model, the faster it forgets. 🧠💨
A few months ago, we fine-tuned a 70B-parameter LLaMA model. It didn’t just forget the new task—it lost the skills and safety mechanisms we worked so hard to build. 😱

Why? That’s when I discovered 𝗖𝗮𝘁𝗮𝘀𝘁𝗿𝗼𝗽𝗵𝗶𝗰 𝗙𝗼𝗿𝗴𝗲𝘁𝘁𝗶𝗻𝗴 and 𝘁𝗵𝗲 𝗦𝗰𝗮𝗹𝗲 𝗟𝗮𝘄.

𝗪𝗵𝗮𝘁’𝘀 𝗛𝗮𝗽𝗽𝗲𝗻𝗶𝗻𝗴? 🤔
Fine-tuning large models often causes catastrophic forgetting. Larger models (70B+) forget faster than smaller ones (7B).

Why?
 • High-Dimensional Loss Landscape 🌄: Bigger models have more catastrophic basins. Even small shifts can push them into these basins, causing knowledge loss.
 • Ripple Effects of Fine-Tuning 🌊: Tuning one layer can trigger a chain reaction, destroying unrelated skills.

𝗧𝗵𝗲 𝗦𝗰𝗮𝗹𝗲 𝗟𝗮𝘄 𝗘𝘅𝗽𝗹𝗮𝗶𝗻𝗲𝗱 📊
A 13B model lost double the capabilities of a 7B on unrelated tasks after fine-tuning.
Bigger models = Bigger forgetfulness risks ⚠️

𝗛𝗶𝘀𝘁𝗼𝗿𝗶𝗰𝗮𝗹 𝗧𝗶𝗺𝗲𝗹𝗶𝗻𝗲 (𝗪𝗵𝗲𝗻 𝗶𝘁 𝘀𝘁𝗮𝗿𝘁𝗲𝗱)

 • 1989: "Catastrophic Forgetting" coined by McCloskey & Cohen.
 • 2016: French et al. confirm neural networks suffer from catastrophic forgetting.
 • 2017: Kirkpatrick et al. propose Elastic Weight Consolidation (EWC) to mitigate forgetting.
 • 2023: Luo et al. find larger LLMs (1B-7B) forget more severely.
 • 2023: Lermen et al. show fine-tuning impacts LLM adaptability.
 • 2024: Kalajdzievski et al. identify an inverse relationship between fine-tuning and forgetting with LoRA.
 • 2025: Google Research introduces Nested Learning to address catastrophic forgetting in AI.

𝗧𝗵𝗲 𝗟𝗼𝘀𝘀 𝗟𝗮𝗻𝗱𝘀𝗰𝗮𝗽𝗲: 𝗔 𝗛𝗶𝗱𝗱𝗲𝗻 𝗧𝗿𝗮𝗽 🕳️
Imagine a model’s knowledge as a wide canyon. Fine-tune in the wrong direction, and it’s like walking it off a cliff. Even a small update can push it into a "catastrophic basin"—performance collapses. 💣

𝗗𝗼𝗲𝘀 𝗣𝗮𝗿𝗮𝗺𝗲𝘁𝗲𝗿-𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗧𝘂𝗻𝗶𝗻𝗴 𝗛𝗲𝗹𝗽? 🚫
LoRA freezes most parameters, but even minimal updates can lead to massive skill loss. Fine-tuning GPT-3.5 on just 10 adversarial examples caused an 88.8% increase in harmful responses. 💀

𝗪𝗵𝗮𝘁’𝘀 𝘁𝗵𝗲 𝗦𝗼𝗹𝘂𝘁𝗶𝗼𝗻? 𝗚𝗲𝗼𝗺𝗲𝘁𝗿𝘆 🔄
Forget keeping weights close—preserve the geometry of the loss landscape.
 Methods like Sharpness-Aware Minimization (SAM) and Functionally Invariant Paths (FIP) reshape the landscape for stability. 🧭

𝗧𝗵𝗲 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆 🔑

 • Bigger models = Bigger forgetfulness risks ⚠️
 • Small updates can lead to catastrophic forgetting 💥
 • Flat, wide loss landscapes are your best defense 🌍

The real challenge is optimizing the loss landscape geometry to prevent disastrous forgetting. 🌟

♻ Repost if you’re building or fine-tuning LLMs—your team needs to know! 🚀
P.S. Ever destroyed a model with just one fine-tune? Share your story in the comments! 🗣️👇

The Turing Test isn’t broken—it’s just irrelevant now.

AI can sound human without understanding a thing.

Mimicry is easy; reasoning isn’t.

Old benchmarks can’t measure modern capabilities.

We need tests that track reliability and real problem-solving.

Impressing humans for a few minutes isn’t intelligence.

This meme captures the gap perfectly.

taplio

Ashish Patel 🇮🇳's Linkedin Analytics

Ashish Patel 🇮🇳

Performance ScoreAshish Patel 🇮🇳's LinkedIn Performance

Ashish Patel 🇮🇳's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?

taplio

.css-6ytdur{background-image:linear-gradient(284deg, rgba(39, 43, 72, 1) 0%, rgba(33, 47, 183, 1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Ashish Patel 🇮🇳's Linkedin Analytics

Ashish Patel 🇮🇳

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Ashish Patel 🇮🇳's LinkedIn Performance

Ashish Patel 🇮🇳's Best Posts .css-1emoo84{font-size:var(--chakra-fontSizes-md);font-weight:400;color:var(--chakra-colors-gray-600);}(last 30 days)

Ready to grow your LinkedIn presence?

Ashish Patel 🇮🇳's Linkedin Analytics

Performance ScoreAshish Patel 🇮🇳's LinkedIn Performance

Ashish Patel 🇮🇳's Best Posts (last 30 days)