Tirthankar Ghosal (@TirthankarSlg) header
Tirthankar Ghosal (@TirthankarSlg) profile picture
Knoxville, TN
Joined January 2017
Photo by Tirthankar Ghosal (@TirthankarSlg)Photo by Tirthankar Ghosal (@TirthankarSlg)Photo by Tirthankar Ghosal (@TirthankarSlg)Photo by Tirthankar Ghosal (@TirthankarSlg)Photo by Tirthankar Ghosal (@TirthankarSlg)Photo by Tirthankar Ghosal (@TirthankarSlg)Photo by Tirthankar Ghosal (@TirthankarSlg)Photo by Tirthankar Ghosal (@TirthankarSlg)Photo by Tirthankar Ghosal (@TirthankarSlg)Photo by Tirthankar Ghosal (@TirthankarSlg)
Tirthankar Ghosal retweeted
😊 I also just uploaded a new version of the lecture notes. cs229.stanford.edu/main_note… (it's still, and perhaps will always be work-in-progress, given the field is moving fast. Any mistakes are mine :) and any feedback is more than welcome :).
Instead of watching 2 hours of Netflix tonight, watch this Stanford lecture. It’s the clearest end-to-end explanation I’ve seen of how ChatGPT and Claude are actually built. From tokenization and BPE all the way to the transformer architecture, training pipeline, and next-token decoding. Whether you’ve never written a line of AI code or you’ve been shipping agents every day, this 2h 34min session will click things into place that most people take years to figure out. Bookmark it and clear the time this weekend, because this might be the single most valuable thing you learn all month.
72
929
17
6,495
567,690
Tirthankar Ghosal retweeted
I couldn't find NeurIPS workshops split by location (Sydney, Paris, Atlanta), so here you go: samgijsen.github.io/neurips-… Sattelite locations have more workshops than I expected (30 and 26 respectively)
2
17
110
17,097
Tirthankar Ghosal retweeted
现在很多人聊 Research Agent,默认期待是: 读论文、找 gap、想 idea、做实验、写 paper。 但耶鲁大学的这篇论文问了一个更深入的问题: LLM 生成的研究想法,和人类研究者真正做出来的 paper idea,差距到底在哪里? 论文的核心概念叫 research taste,可以理解成「研究品味」:你通常会发现什么样的问题,用什么方式把它变成一个 contribution。 它的实验设计挺巧。 作者从真实论文里抽取 human idea,再反推 4 到 8 篇最可能启发这个 idea 的 prior works。然后把这些 prior works 的标题和摘要给 LLM,让模型在同样的文献上下文里生成一个新的 research idea。 这样比较的重点,就从「单个 idea 看起来新不新」推进到「一批 idea 的分布像不像人类研究者」。 论文一共用了 11,683 个 human ideas,覆盖 ML 会议论文和 Nature Communications 里的自然科学论文;模型包括 Claude、Gemini、GPT、Qwen、DeepSeek 等 9 个设置。 结果很有意思:人类 idea 的分布明显更宽。 在人类论文里,只有 12.1% 的 idea 属于 bridge opportunity,也就是把不同文献、方法或证据流连接起来。 但在 LLM 生成的 idea 里,这个比例变成了 47.1% 到 64.2%。 方法层面也类似。 人类论文里,synthesis / unification 只占 5.1%;LLM 这里是 22.5% 到 38.7%。 简单说,模型很喜欢把研究问题理解成: 这里有两个东西,可以把它们整合一下。 论文还发现,给模型更多上下文也没有明显解决这个问题。full-paper context 版本没有让分布更接近人类;开启 thinking mode 甚至可能让模型更强化这种 bridge-and-synthesis 倾向。 这篇最有意思的地方在于,它没有简单说 LLM idea 好或不好,而是指出了一个更隐蔽的问题: LLM 可以生成很多看起来合理的研究想法,但这些想法的类型可能高度集中。 这对 AI for Science 和 Research Agent 很关键。 对于真正强的 Research Agent,难点在于能不能形成更宽的 problem-finding 能力:发现 failure、识别机制、提出测量工具、构造系统、修改局部假设,而不总是回到「整合两个已有东西」这条安全路径。 自动科研如果缺少这种多样性,最后可能会变成一种很流畅的同质化机器。 📎 arxiv: arxiv.org/abs/2607.01233
37
182
8
794
79,080
Tirthankar Ghosal retweeted
🌉#NAACL2027 will be June 1-5 in San Francisco, CA!🌉 ARR submission deadline: Oct. 12, 2026
2
26
1
208
16,285
Tirthankar Ghosal retweeted
probably the best blog i have read for some time viewing SFT, RL, and OPD as different ways of reshaping a model's distribution makes their tradeoffs super intuitive. - SFT pulls toward a fixed external target - RL moves along the reward gradient on on-policy samples - OPD sits in between, using a teacher signal but on student-generated data, which is why it inherits RL's anti-forgetting properties even when the teacher itself was an overtrained SFT model. the post is heavily grounded in recent literature and uses the distributional perspective as a unifying bridge across all three paradigms, i really like the point it argues the load-bearing ingredient is on-policy data and OPD's convergence to RL-like outcomes is the strongest evidence
11
210
7
1,827
102,413
Tirthankar Ghosal retweeted
Resources for grant writing with AI github.com/eseckel/ai-for-gr…
13
155
6
1,214
128,891
Tirthankar Ghosal retweeted
Replying to @WiNLPWorkshop
@WiNLPWorkshop is partnering with the @aaclmeeting D&I Committee to launch a mentorship program for students, early-career researchers, and first-time attendees! 🤝 tinyurl.com/2du9459n No one should navigate NLP alone — let’s build a community where everyone belongs 💙🌍
2
1
130
Tirthankar Ghosal retweeted
Hugging Face just dropped a 200 page playbook on training LLMs. It covers everything from pre-training to post-training and infrastructure with real examples of what worked and what didn't. 100% free and Opensource.
10
26
160
10,465
Tirthankar Ghosal retweeted
Seriously... why did no one tell me about this?! BrowserOS is a 100% open-source agentic browser alternative to ChatGPT Atlas and Perplexity Comet 🤯 Gave it a spin, it’s surprisingly smooth and stable. Repo in 🧵 ↓
13
14
2
64
9,538
Agentic RAG Tech Stack
20
203
4
1,121
68,417
Tirthankar Ghosal retweeted
𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚 and what you need to know about it as an AI Engineer? Simple naive RAG systems are rarely used in real world applications. We are usually adding some agency to the RAG system - ideally a minimal amount. There is 𝗻𝗼 𝘀𝗶𝗻𝗴𝗹𝗲 𝗯𝗹𝘂𝗲𝗽𝗿𝗶𝗻𝘁 on how the RAG system needs to be extended to solve your specific business use case, you will need to adapt. For this you need to understand the potential moving pieces in Agentic RAG, let’s explore some of them: 𝟭. Analysis of the user query: we pass the original user query to a LLM based Agent for analysis. This is where: ➡️ The original query can be rewritten, sometimes multiple times to create either a single or multiple queries to be passed down the pipeline. ➡️ The agent decides if additional data sources are required to answer the query. 𝟮. If additional data is required, the Retrieval step is triggered. In Agentic RAG case, we could have a single or multiple agents responsible for figuring out what data sources should be tapped into, few examples: ➡️ Real time user data. This is a pretty cool concept as we might have some real time information like the current location of the user. ➡️ Internal documents that a user might be interested in. ➡️ Data available on the web. ➡️ … 𝟯. Retrieved data is consolidated and Reranked by a more powerful model compared to regular embedder. Data points are significantly narrowed down. 𝟰. If there is no need for additional data, we try to compose the answer (or multiple answers or a set of actions) straight via an LLM. 𝟱. The answer gets analysed, summarised and evaluated for correctness and relevance: ➡️ If the Agent decides that the answer is good enough, it gets returned to the user. ➡️ If the Agent decides that the answer needs improvement, we try to rewrite the user query and repeat the generation loop (this number of times the loop could run should be capped at a certain number). 𝗠𝘆 𝗛𝗼𝗻𝗲𝘀𝘁 𝗧𝗵𝗼𝘂𝗴𝗵𝘁𝘀: ✅ Simplify wherever possible, you don’t need most of the pipeline in most of the cases. ✅ Focus on proper preprocessing of the data and the Reranking step. It is enough for 90%+ use cases. ❗️ Agentic Systems go far beyond RAG in real enterprises. It becomes more interesting when you move beyond question answering machines to systems that can automate actions after routing the initial human intent appropriately. More on this in future posts, stay tuned! What are your thoughts on Agentic RAG? Let me know in the comments! 👇 #LLM #AI #MachineLearning
12
96
3
567
28,485
I finally understand the fundamentals of building real AI agents. This new paper “Fundamentals of Building Autonomous LLM Agents” breaks it down so clearly it feels like a blueprint for digital minds. Turns out, true autonomy isn’t about bigger models. It’s about giving an LLM the 4 pillars of cognition: • Perception: Seeing and understanding its environment. • Reasoning: Planning, reflecting, and adapting. • Memory: Remembering wins, failures, and context over time. • Action: Executing real tasks through APIs, tools, and GUIs. Once you connect these systems, an agent stops being reactive it starts thinking. Comment "Paper" and I'll DM you the link.
472
350
22
2,227
210,476
Tirthankar Ghosal retweeted
What is RAG? What is Agentic RAG? 𝐑𝐀𝐆 (𝐑𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐥-𝐀𝐮𝐠𝐦𝐞𝐧𝐭𝐞𝐝 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐨𝐧) RAG connects a generation model to external knowledge through retrieval. Here’s how it works - 1./ A user submits a query. 2./ The system searches a pre-indexed set of documents (typically stored in a vector database). 3./ The most relevant chunks are retrieved. 4./ These chunks are appended to the original query. 5./ The combined input is sent for generation. The goal? To provide the model with context so it can generate more accurate, source-aware responses. But in traditional RAG, everything happens in a single pass - no planning, no evaluation, no retrying. ------ 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐑𝐀𝐆 Agentic RAG builds on the same foundation but introduces intelligent agents, each with a specific role - to improve the overall process. Instead of a single static pipeline, Agentic RAG becomes 𝐚 𝐦𝐮𝐥𝐭𝐢-𝐬𝐭𝐞𝐩, 𝐚𝐝𝐚𝐩𝐭𝐢𝐯𝐞 𝐬𝐲𝐬𝐭𝐞𝐦. Here's what it typically includes - 1./ A Planning Agent that breaks down the user query and decides what needs to be retrieved. 2./ A Retrieval Agent that reformulates the query (if needed) and pulls in information, not just from documents, but potentially from APIs, tools, or dynamic sources. 3./ A Generator Agent that constructs a response using the retrieved data. 4./ A Judge or Evaluation Agent that reviews the output. If it’s not good enough, the system can refine its plan or regenerate the answer. This setup allows for - - Iterative reasoning - Self-correction - Tool integration - Dynamic, context-aware responses ------------ Follow @techNmak
58
216
3
1,201
79,431
Tirthankar Ghosal retweeted
Combining the benefits of RL and SFT with on-policy distillation, a promising approach for training small models for domain performance and continual learning.
Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other approaches for a fraction of the cost. thinkingmachines.ai/blog/on-…
99
213
32
2,794
488,700
Tirthankar Ghosal retweeted
Writing a lit review is easy if you have a strategy. Modern AI tools can also make it fast and efficient. Here is how:
1
7
44
2,882
Tirthankar Ghosal retweeted
Build MCP AI Agents with reasoning, system prompts, and tool orchestration. Nanobot wraps existing MCP servers into intelligent agents and renders React components directly in chat via MCP-UI. 100% open-source.
11
68
2
418
28,560
Google open-sourced LangExtract Python library! It uses LLMs to extract entities, attributes, and relations with exact source grounding from unstructured documents. Flexible LLM support (Gemini, OpenAI, Ollama) 100% open-source.
15
181
6
1,104
73,951
Tirthankar Ghosal retweeted
Build an MCP server that extract structured data from complex documents!
Let’s build an MCP server that can extract structured data from visually complex documents using Agentic Document Extraction (100% local).
6
86
1
483
47,306
Tirthankar Ghosal retweeted
Toolkit for fine-tuning and training large language and vision models
3
37
1
229
12,477
Tirthankar Ghosal retweeted
Stanford CS229: Building Large Language Models This brilliant 1.5h lecture unpacks how ChatGPT-like models are built: From tokenization & scaling laws to training hurdles, benchmarks, SFT/RLHF, and efficiency Lecture link in 🧵 ↓
13
58
394
40,162