Bhargav Bhikkaji (@bbhikkaji) profile picture

Founder @ https://nitter.gq/t.co/W5bf3bNl9Z

San Jose
Joined June 2008
Photo by Bhargav Bhikkaji (@bbhikkaji)
It’s unbelievable that Vaibhav Suryavanshi is already being compared by legends of the game to the likes of Viv Richards, Sachin Tendulkar, Brian Lara and even beyond to likes of Don Bradman. This kid hasn’t played a single official international match yet.
35
Bhargav Bhikkaji retweeted
Most presenters lose their audience before the first slide even appears. After studying hundreds of presentation openings, here are five that consistently work: 1. Start with a “Did you know…” Shock them with a fascinating fact or unexpected insight. Curiosity drives engagement. 2. Use a “What if…” Invite your audience into a possible future. This makes your idea feel like an adventure, not a lecture. 3. Share a secret. Everyone loves insider knowledge. Phrases like “They don’t want you to know this…” spark instant intrigue. 4. Name your audience. “This one’s for designers who…” When people hear themselves, they listen. 5. Address a struggle. “If you’ve ever felt stuck trying to…” connects you emotionally with those who face that exact challenge. The pattern behind all great hooks? They promise transformation. When people sense that something valuable, surprising, or emotional is about to unfold, they keep listening. Don't assume your audience will listen just because you're speaking. Use one of these five openers to earn their attention.
19
209
12
1,416
170,679
Bhargav Bhikkaji retweeted
The paper shows OpenAI’s smaller GPT-OSS 20B beats the larger 120B on key tasks. It also needs 5x less GPU memory and 2.6x lower energy per response than 120B. The core question is simple, do sparse mixture of experts models actually get better just by getting bigger. Mixture of experts means only a few expert blocks activate for each token, which reduces compute while keeping skill. The authors compare 8 open models across 10 benchmarks under the same inference setup, then validate differences with standard statistical tests. Results put both GPT-OSS models in the middle of the pack, with clear strength in code generation and a clear weakness on Chinese tasks. Surprisingly, 20B tops 120B on MMLU and HumanEval and stays more concise, which signals routing or training inefficiency in the larger variant. For practical picks, 20B delivers better cost performance, higher throughput, and lower latency, so it suits English coding and structured reasoning more than multilingual or professional domains. ---- Paper – arxiv. org/abs/2508.12461 Paper Title: "Is GPT-OSS Good? A Comprehensive Evaluation of OpenAI's Latest Open Source Models"
8
36
4
188
12,321
Bhargav Bhikkaji retweeted
Peter Thiel on what it takes to build monopoly-style businesses.
4
23
173
16,174
7 Layers of LLM Stack
14
394
7
2,096
193,282
Bhargav Bhikkaji retweeted
👉Here’s a real CI/CD flow most companies follow: 1.Code pushed → Git triggers pipeline 2. Build → Dockerize app 3.Test → Unit & Integration 4.Deploy → Staging → Manual Approval → Production 5.Monitor → Grafana + Prometheus 6.Rollback → Blue-Green or Canary. Yes, in most companies, manual approval is required before deploying to production.
1
62
369
25,110
Bhargav Bhikkaji retweeted
"LLMs + Coding Agents = Security Nightmare" The title says it all. The article goes on to explain how these problems are fundamental. I posted about now I use LLMs last week (basically to teach myself new topics, and I ask multiple LLMs to review each other's output too). I personally do not use AI coding agents and I have given the same advice to our developers to be very cautious on them. Our stance is that LLM generated code should never get into production without full scale human review, particularly security review and once you include those reviews, productivity gains are elusive. garymarcus.substack.com/p/ll…
44
116
8
807
54,540
Bhargav Bhikkaji retweeted
OpenAI hasn’t open-sourced a base model since GPT-2 in 2019. they recently released GPT-OSS, which is reasoning-only... or is it? turns out that underneath the surface, there is still a strong base model. so we extracted it. introducing gpt-oss-20b-base 🧵
160
424
111
5,988
930,399
Bhargav Bhikkaji retweeted
RAG can skip pre-built graphs by building a small, per-question reasoning graph on the fly. LLMs struggle when answers need many facts stitched together. Graph-based RAG tries to fix this by first turning the whole corpus into a big graph, which is slow, costly, and often mismatched to the question’s logic. This paper proposes LogicRAG, a runtime planner. It breaks the user query into subproblems, then makes a directed acyclic graph that shows which subproblem depends on which. It orders that graph with a topological sort, then solves subproblems in that order, retrieving only what each step needs and feeding the results forward. To keep tokens under control, it keeps a rolling memory, a short running summary of the facts found so far, and prunes context that does not help the next step. It also merges subproblems that sit at the same level of the graph to avoid repeated retrievals. And it samples subproblems without replacement so the model does not stall by asking near-duplicate subqueries. Across 3 multi-hop QA sets, LogicRAG beats strong GraphRAG systems without any offline graph building. On 2WikiMQA it reaches 64.7% string accuracy and 62.5% LLM accuracy, a +14.7% jump over the next best baseline. On HotpotQA it gets 54.8% and 62.6%, and on MuSiQue it gets 30.4% and 37.5%. Query-time cost is practical too since there is no graph construction stage. On 2WikiMQA the average runtime is 9.83 s with 1,777.9 tokens, while several graph-heavy methods take more time or use far more tokens. Net effect, retrieval follows the question’s logic instead of a fixed, pre-built map, so answers are more accurate and cheaper to get. ---- Paper – arxiv. org/abs/2508.06105 Paper Title: "You Don't Need Pre-built Graphs for RAG: Retrieval Augmented Generation with Adaptive Reasoning Structures"
9
41
3
220
16,365
Bhargav Bhikkaji retweeted
Protect your children...
1,021
24,308
920
83,875
15,108,596
Bhargav Bhikkaji retweeted
Is Chain-of-Thought Reasoning of LLMs a Mirage? ... Our results reveal that CoT reasoning is a brittle mirage that vanishes when it is pushed beyond training distributions. This work offers a deeper understanding of why and when CoT reasoning fails, emphasizing the ongoing challenge of achieving genuine and generalizable reasoning. ... Our findings reveal that CoT reasoning works effectively when applied to in-distribution or near in-distribution data but becomes fragile and prone to failure even under moderate distribution shifts. In some cases, LLMs generate fluent yet logically inconsistent reasoning steps. The results suggest that what appears to be structured reasoning can be a mirage, emerging from memorized or interpolated patterns in the training data rather than logical inference. ... Together, these findings suggest that LLMs are not principled reasoners but rather sophisticated simulators of reasoning-like text.
193
886
204
5,719
794,211
Bhargav Bhikkaji retweeted
The survey shows that most agent projects still talk past each other, so tasks get stuck or repeated. It lines up 24 frameworks side by side and argues that a common spec would make them cooperate instead of collide. An agent workflow is just the playbook that tells a language model which sub tasks to run, in what order, and with which outside tools. Without that playbook, a single prompt either wastes tokens guessing next steps or forgets what it did 2 minutes ago. The authors group systems along 2 axes. Function covers skills like task planning, tool calls, memory, and team work, while architecture tracks how they wire roles, data flow, and protocols. They find broad agreement on using python chains and JSON configs, yet every lab invents its own tool schema, message format, and role names. That lack of harmony blocks cross platform jobs and makes security bugs such as tool poisoning harder to fix. The paper points to early fixes, for example open protocols like MCP and A2A, planner‑executor loops that cut token spend, and scheduling tricks that prevent agents fighting over the same resource. Still, until the community writes a shared grammar for workflows, builders will keep shipping one‑off hubs that cannot plug into each other. ---- Paper – arxiv. org/abs/2508.01186 Paper Title: "A Survey on Agent Workflow -- Status and Future"
2
31
2
173
11,729
Bhargav Bhikkaji retweeted
NVIDIA's brilliant paper gives a lot of details and actionable techniques. 🎯 Small Language Models (SLMs) not LLMs are the real future of agentic AI. They gave a recipe for swapping out the large models with SLMs without breaking anything and show that 40%‑70% of calls in open agents could switch today. It argues that SLMs, already match big models on many routine agent tasks, cost far less to run, and slot neatly into mixed‑model pipelines, so most agent calls should shift to SLMs. 🗝️ The central claim SLMs that fit on a laptop already handle the narrow language chores inside most agents. Because those chores rarely need open‑ended chat, the authors say an SLM‑first design is the natural default, and large models become occasional helpers. @NVIDIAAIDev 👏
21
197
12
1,055
78,878
Bhargav Bhikkaji retweeted
NVIDIA just dropped paper exposing a $57 billion AI industry mistake. While Big Tech keeps pushing expensive LLMs like ChatGPT & Claude... Small language models handle 70% of AI agent work at 1/30th the cost. Here's why this changes everything: (hint: less is more)
Readers added context they thought people might want to know
This post has many false claims, and it's mostly just hyping this paper: arxiv.org/pdf/2506.02153 "Exposing a $57 billion AI industry mistake" is just a random, made-up statement. "Small language models handle 70% of AI agent work at 1/30th the cost." This statement is also unsupported.
58
231
22
1,302
199,387
Bhargav Bhikkaji retweeted
what are large language models actually doing? i read the 2025 textbook "Foundations of Large Language Models" by tong xiao and jingbo zhu and for the first time, i truly understood how they work. here’s everything you need to know about llms in 3 minutes↓
81
887
39
6,879
1,118,225
Bhargav Bhikkaji retweeted
Slides for my lecture “LLM Reasoning” at Stanford CS 25: dennyzhou.github.io/LLM-Reas… Key points: 1. Reasoning in LLMs simply means generating a sequence of intermediate tokens before producing the final answer. Whether this resembles human reasoning is irrelevant. The crucial insight is that transformer models can become nearly arbitrarily powerful by generating many intermediate tokens, without the need of scaling the model size (arxiv.org/abs/2402.12875). 2. Pretrained models, even without any fine-tuning, are capable of reasoning. The challenge is that reasoning-based outputs often don’t appear at the top of the output distribution, so standard greedy decoding fails to surface them (arxiv.org/abs/2402.10200) 3. Prompting techniques (e.g., chain-of-thought prompting or "let’s think step by step") and supervised finetuning were commonly used to elicit reasoning. Now, RL finetuning has emerged as the most powerful method. This trick was independently discovered by several labs. At Google, credit goes to Jonathan Lai on my team. Based on our theory ( see point 1), scaling RL should focus on generating long responses rather than something else. 4. LLM reasoning can be hugely improved by generating multiple responses and then aggregating them, rather than relying on a single response (arxiv.org/abs/2203.11171).
62
483
44
3,100
470,763
Bhargav Bhikkaji retweeted
There will be more jobs in AI that we have yet to imagine!
19
320
17
2,229
256,151