โ† Back to leaderboard
techwith_ram

๐—ฟ๐—ฎ๐—บ๐—ฎ๐—ธ๐—ฟ๐˜‚๐˜€๐—ต๐—ป๐—ฎโ€” ๐—ฒ/๐—ฎ๐—ฐ๐—ฐ

Verifiedactive
@techwith_ram

Sr. DS โ€ข AI Updates โ€ข Latest news & AI tools โ€ข Views are my own

RosterAI Model News & ReleasesAI Agents & Productivity Tools
12,997 followers ยท 1,014 following ยท Last polled 3h ago
Perf. Score
43
Rank #288
Value Score
35.7
#181 of 273 priced
Median views
569
p25 503 ยท trailing 7d
Followers
13K
+0.1% 7d
ER (impr.)
4.78%
eng รท impressions
ER (foll.)
0.72%
eng รท followers
Est. CPM
$87.87
QT-rate basis
WoW views
+11.9%
flat

Rates & value ยท est. from 569 median organic views

overpriced
Quote tweetbasis
$50
$87.87 / 1K views
Post
โ€”
no CPM
Thread
โ€”
no CPM
Retweet
โ€”
no CPM
Value Score 35.7 ยท rank #181 of 273 priced$2631.58 / 1K engagementsrate age unknown (pre-audit import)

Trends

Follower growth
Median views / day
Engagement rate
View distribution ยท trailing-7d posts

Each blue dot is one organic post. The Performance Score ranks by the median (569), not the mean (2.0K) โ€” so a single viral post doesn't inflate a creator's reach. The p25 floor (503) is the level 75% of posts clear.

Recent posts ยท ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
Transformer Architecture Explained in 3 ways here: - Encoder only - Deconder only - Encoder - Decoder both https://t.co/PzerDRxN0qfrozen20d ago19K43492204485.10%
NVIDIA released NV-Reason-CT A 3D medical AI model that lets an LLM reason over 13,824 visual tokens from a single CT. - Language backbone: Qwen3.5-4B - 3D vision encoder: Primus, initialized from COLIPRI - Input: one-channel chest or abdominal CT in .nii or .nii.gz format - Preprocessing: LPS orientation, 2 mm isotropic resampling, and an anatomy-aware 192 x 192 x 192 voxel crop - Patch grid: 24 x 24 x 24 from non-overlapping 8 x 8 x 8 patches - Visual interface: all 13,824 projected visual tokens are passed to the language model without spatial merging - Position encoding: depth, height, and width are represented with 3D MRoPE Github: https://t.co/tB1zrKCe7f Huggingface: https://t.co/ToND2QRRaW26d ago11K201292
2
216
4.19%
RL Guide by Unsloth Learn all about Reinforcement Learning (RL) and how to train your own DeepSeek-R1 reasoning model with Unsloth using GRPO. A complete guide from beginner to advanced. Link: https://t.co/gKMLIuxWDH https://t.co/IN2sxorsdifrozen27d ago10K25837303266.16%
๐—ฆ๐—ฒ๐—น๐—ณ-๐—˜๐˜ƒ๐—ผ๐—น๐˜ƒ๐—ถ๐—ป๐—ด ๐—”๐—ด๐—ฒ๐—ป๐˜ ๐—›๐—ฎ๐—ฟ๐—ป๐—ฒ๐˜€๐˜€ The agents donโ€™t fail loudly. They fail the same way, again and again, and nobody notices until a user does. So the question I keep coming back to isnโ€™t โ€œhow smart is the agent?โ€ Itโ€™s โ€œhow does it get better without getting worse?โ€ This harness answers that. The agent perceives, thinks, plans and acts like any other. The difference is what happens after a mistake. It doesnโ€™t just record that something failed. It compares the failed run against a good one and finds the exact step where things went wrong. It doesnโ€™t patch one-off errors either. It groups failures by root cause, so it fixes the pattern, not the symptom. No hardcoded answers. No sneaking test data into the prompt. And nothing changes on trust alone. Every update the agent proposes for itself has to pass a quality gate and a performance gate before itโ€™s allowed in. If it makes things worse, itโ€™s rejected. That last part is what matters most. Self-improvement without protection is just drift with extra steps. Precise diagnosis. Real generalization. Regression protection. Validated updates.2d ago8.5K22643602746.43%
I keep seeing people call an LLM an โ€œAI agentโ€ just because it can use a tool. But the LLM is only one part of the system. The model does the reasoning and decides what needs to happen. The agent harness is what actually connects everything together. It manages tools, permissions, context, memory, sandboxing, sub-agents, and the loop that keeps the agent working until the task is done. Think of it like this: - Model = decides - Harness = executes So, what happens between those two? A tool gets called. The result comes back. The context gets updated. Permissions are checked. Memory might be used. Another tool might be called. And the loop continues until the agent reaches the goal.frozen18d ago7.4K17524502035.52%
Ollama vs vLLM vs SGLang Ollama, vLLM and SGLang all serve LLMs. Theyโ€™re built for very different jobs. Ollama assumes the bottleneck is access. It packages a quantized model (GGUF through llama.cpp, or MLX on Apple Silicon) into a single local runtime. By default it serves requests in order and runs entirely offline. Itโ€™s built for one person at one machine. vLLM assumes the bottleneck is GPU memory & idle time. Continuous batching lets new requests join the running batch at every decoding step instead of waiting for the slowest sequence to finish. PagedAttention borrows from operating-system virtual memory: the KV cache is split into fixed-size pages, so memory is allocated on demand rather than reserved up front. The result is high throughput across thousands of unrelated prompts. SGLang assumes the bottleneck is repetition. Agents and multi-turn conversations send the same long system prompts, tool definitions and chat history again and again. RadixAttention stores those prefixes in a radix tree so shared context is computed once and reused, and a cache-aware router sends each request to the worker that already holds its prefix. Constrained decoding then forces the output into valid JSON or a regex pattern. Three engines, three assumptions about where compute is lost: Access โ†’ Ollama Concurrency โ†’ vLLM Shared context โ†’ SGLangfrozen11d ago7.3K219401411866.34%
NVIDIA is turning agent skills into RL environments. Skill2Env takes public SKILL.md files, turns them into executable tasks, and uses them to train agents through reinforcement learning The best part is 7,971 tasks across 13 domains. After 300 steps of outcome-only RL, Qwen3.8-27B improved from 49.4% to 54.1% on Terminal-Bench 2.1 and from 33.4% to 37.7% pass@1 on S2EBench, their hand-verified held-out benchmark Paper: https://t.co/FBxjswPRbP Check this site. Where you can chat with the paper itself: https://t.co/qBb2uMrSfR Source: Dair. aifrozen26d ago5.9K127201321254.85%
๐Ÿฑ ๐—ฅ๐—”๐—š ๐—”๐—ฟ๐—ฐ๐—ต๐—ถ๐˜๐—ฒ๐—ฐ๐˜๐˜‚๐—ฟ๐—ฒ๐˜€ Most of the RAG systems fail in the same place. It isnโ€™t the model. Itโ€™s how the context gets found. ๐—•๐—ฎ๐˜€๐—ถ๐—ฐ ๐—ฅ๐—”๐—š embeds your documents, finds the closest chunks and hopes they hold the answer. That works for simple questions. Production systems need more, and five architectures have emerged to fill the gap. ๐—›๐˜†๐—ฏ๐—ฟ๐—ถ๐—ฑ ๐—ฅ๐—”๐—š runs two searches at once. Dense vectors capture meaning, while a BM25 keyword index catches exact terms that embeddings tend to miss: product codes, names, acronyms. Reciprocal rank fusion merges the two ranked lists into one. ๐—š๐—ฟ๐—ฎ๐—ฝ๐—ต๐—ฅ๐—”๐—š stops treating knowledge as isolated chunks. It extracts entities and the relationships between them into a knowledge graph, then retrieves the relevant subgraph and summaries of related clusters. Itโ€™s built for questions whose answers span several connected facts. ๐—”๐—ด๐—ฒ๐—ป๐˜๐—ถ๐—ฐ ๐—ฅ๐—”๐—š turns retrieval into a plan. A planner agent decides which tools to use (vector search, web search, a SQL database), and a reasoner agent combines the results. If the evidence is weak, it loops back and searches again. ๐—–๐—ผ๐—ฟ๐—ฟ๐—ฒ๐—ฐ๐˜๐—ถ๐˜ƒ๐—ฒ ๐—ฅ๐—”๐—š questions its own retrieval. An evaluator grades what came back before anything reaches the model. Relevant results go through, irrelevant ones trigger a web search, and uncertain ones send the query back to be rewritten. ๐— ๐˜‚๐—น๐˜๐—ถ๐—บ๐—ผ๐—ฑ๐—ฎ๐—น ๐—ฅ๐—”๐—š treats a document as more than text. Text, images, charts and tables are embedded into one shared index using models such as CLIP or ColPali, so a single query can retrieve a paragraph, a diagram or a scanned table. Each architecture fixes a different failure: Missed keywords โ†’ Hybrid Missing connections โ†’ GraphRAG Questions that need several sources โ†’ Agentic Unreliable retrieval โ†’ Corrective Knowledge locked in visuals โ†’ Multimodalfrozen10d ago4.9K15322211526.73%
The Jev and laya story is interesting. If you don't know read this. An American company launched a model that was insanely good at zero-shot classification. Twitter was full of it. Launch videos, demos, influencers reposting the same thing, everyone talking about this โ€œnewโ€ approach. And honestly, the ecosystem around the launch was really good. But somewhere in Kerala, an Indian researcher named Nandan Kishore was watching all of this and thought: โ€œWaitโ€ฆ I built something very similar more than a year ago.โ€ He had already published the code, datasets and benchmarks. But instead of people looking at the work, the first question was: โ€œOkay, but can you actually prove that your model is good?โ€ So he didn't argue. He decided to build something crazy. And then he went back to work. What came out was Jev, a 429M-parameter model built around the same idea, released openly with the code and research behind it. And suddenly, people started testing it. The crazy part? This isn't just another AI launch story. It's a reminder that talent is everywhere. India has an insane amount of talent. Sometimes the problem isn't talent. It's visibility. It's distribution. It's getting heard in a world where attention is extremely crowded. But that's exactly why open source matters. You don't need permission to build. You don't need permission to publish. You don't need permission to prove yourself. Life is a game. And maybe we're playing it on hard mode. Here is his opensource project link: https://t.co/rS3CtHG2kl1mo ago3.3K591053202.91%
https://t.co/MEv4XHd6skfrozen7d ago2.9K12211161.09%
Contrastive Language Models A System One Model for Fast and Generalizable Decision-Making from NVIDIA + Stanford This is pre-trained on 60M Nemotron Q&A pairs, mid-trained on 30M synthetic hard negatives, and post-trained on 1M agentic trajectories. Its performance comparable to Jev across computer-use, gaming, and tool-calling tasks, while achieving up to 9ร— lower latency Github Repo: https://t.co/CU8TPOa6bb Blog: https://t.co/AzJmODYWiXfrozen25d ago2.5K43810483.99%
I keep seeing the same pattern with RAG projects. Someone builds a prototype over a weekend. They connect a vector database, add a few PDFs, write a prompt, and ask it ten questions. Every answer looks right, the demo goes well, and everyone is happy. Then it goes to real users, and within a week the complaints start. โ€œIt gave me the old refund policy.โ€ โ€œIt couldnโ€™t find the error code I searched for.โ€ โ€œIt answered confidently, but the answer was completely made up.โ€ The first reaction is usually to blame the model and switch to a bigger LLM. Most of the time that changes nothing, because the model was never the problem. The real issue sits before the model ever sees the question. The right chunk was never retrieved, or it was cut in half during chunking, or it was buried under four passages that only looked relevant. The LLM did exactly what it was told and answered from bad context. Reliable RAG is not about a smarter model. Itโ€™s about giving the model the right information and keeping it honest about what that information says. Hereโ€™s the pipeline Iโ€™d build today: Chunk by structure : Split on headings, not token counts, and add the section heading to every chunk so it keeps its context. Use hybrid retrieval : Vector search understands meaning but misses exact terms like product codes. BM25 catches those. Run both and merge the results with RRF. Rerank 50 down to 5 : Retrieve wide for recall, then use a cross-encoder to keep only the best matches. If nothing passes the threshold, donโ€™t force an answer. Ground every answer : Tell the model to answer only from the sources, cite each claim, and clearly say when the docs donโ€™t have the answer. Evaluate both halves separately : Measure retrieval and generation on their own. Otherwise youโ€™ll never know which part actually broke. Guard and monitor Filter chunks by user permissions, re-index when documents change, and log every request so real failures become new test cases. None of this is fancy. Itโ€™s just the steady engineering work that turns a nice demo into something people can actually trust.frozen6d ago2.5K511280605.28%
๐—ฅ๐—”๐—š ๐˜ƒ๐˜€ ๐—ข๐—ž๐—™ ๐—˜๐˜…๐—ฝ๐—น๐—ฎ๐—ถ๐—ป๐—ฒ๐—ฑ Ask your RAG chatbot โ€œwhatโ€™s our refund policy?โ€ & it might give you the right answer. Or it might give you last yearโ€™s version, because that chunk happened to score 0.83 instead of 0.81. Thatโ€™s the quiet problem with RAG. It finds whatโ€™s similar, not necessarily whatโ€™s true. Google Cloud released something in June that takes a different approach. Itโ€™s called OKF, the ๐—ข๐—ฝ๐—ฒ๐—ป ๐—ž๐—ป๐—ผ๐˜„๐—น๐—ฒ๐—ฑ๐—ด๐—ฒ ๐—™๐—ผ๐—ฟ๐—บ๐—ฎ๐˜, and itโ€™s almost boring in how simple it is. There are no embeddings and no vector database. Itโ€™s just a folder of Markdown files, one concept per file. Your refund policy lives in policies/refund. md. Your revenue definition lives in metrics/revenue. md. Each file has a small header saying what it is, and files link to each other like a wiki. When an agent needs a fact, it doesnโ€™t search. It opens the index, goes straight to the right file and reads it. If the policy changes, you edit one file and commit it to Git. You get a version history, an owner and a date, and nothing has to be re-embedded. ๐—ฆ๐—ผ ๐—ถ๐˜€ ๐—ฅ๐—”๐—š ๐—ฑ๐—ฒ๐—ฎ๐—ฑ? No. OKF works for the small set of things that must be exactly right: definitions, rules, schemas. Nobody is hand-writing a Markdown file for 50,000 support tickets. Thatโ€™s still RAGโ€™s job. The setup that makes sense to me is both, with a router in front. Known facts go to OKF and everything messy goes to RAG. Search what you canโ€™t curate. Curate what you canโ€™t afford to get wrong.frozen8d ago2.0K43813424.74%
RLHF-Reinforcement Learning From Human Feedback The basic idea is pretty simple: First, teach the model to answer well. Humans provide good examples, and we use them to supervised fine-tune the LLM. Then, teach the model what humans prefer. Humans rank different answers, & those preferences are used to train a reward model. Finally, optimize the LLM toward those preferences. PPO updates the policy so it produces responses that receive higher rewards, while the KL penalty keeps it from drifting too far away from the original SFT model.frozen19d ago1.8K41530314.42%
GPU vs CPU vs TPU CPU - It is built around a small number of powerful cores, & each core has its own control logic, cache, & the ability to make decisions about what to do next. That makes CPUs really good at tasks where the work is sequential, unpredictable, or full of branching. GPU - Instead of having a few powerful cores, it has thousands of smaller compute units designed to work together. They can perform the same operation across huge amounts of data at the same time. Thatโ€™s exactly what makes GPUs so good for AI training, inference, graphics, and other highly parallel workloads. TPUs - are designed specifically around the kind of mathematics that powers neural networks. Their systolic arrays let data move through a grid of compute units, where multiplication and accumulation happen continuously. Instead of constantly going back and forth to memory, the computation flows through the chip.frozen16d ago1.8K58310365.43%
LLM evaluation is where you find out if your model is actually good, or just sounds good. A model can write impressive answers and still fail on accuracy, reasoning, consistency, safety, or hallucinations. Thatโ€™s why evaluating an LLM goes beyond asking, โ€œDoes this answer look good?โ€ You need to test things like: โ€ข Accuracy โ€ข Relevance โ€ข Reasoning โ€ข Factuality โ€ข Hallucination โ€ข Safety โ€ข Consistency โ€ข Instruction following And depending on the use case, you might use benchmarks, human evaluation, LLM-as-a-judge, or custom evaluation datasets. Building an LLM is only half the job. Knowing whether it actually works is the other half.frozen22d ago1.5K36900305.01%
๐—ฉ๐—ฒ๐—ฐ๐˜๐—ผ๐—ฟ ๐—ฅ๐—”๐—š ๐˜ƒ๐˜€ ๐—ฉ๐—ฒ๐—ฐ๐˜๐—ผ๐—ฟ๐—น๐—ฒ๐˜€๐˜€ ๐—ฅ๐—”๐—š ๐˜ƒ๐˜€ ๐—”๐—ด๐—ฒ๐—ป๐˜๐—ถ๐—ฐ ๐—ฅ๐—”๐—š RAG tutorials teach you one pattern: chunk the document, embed it, store the vectors, search by similarity. It works. Until someone asks a question where the answer depends on where something sits in the document, not just what it says. Hereโ€™s how three approaches handle the same problem, and why the field is slowly moving from searching text to navigating structure. ๐—ฉ๐—ฒ๐—ฐ๐˜๐—ผ๐—ฟ ๐—ฅ๐—”๐—š: find what looks similar. You take a long document, split it into fixed-size chunks, turn each chunk into a vector with an embedding model and store them in a vector database. When a question comes in, it gets embedded with the same model, and the system pulls back the top-k chunks that sit closest to it in vector space. The LLM writes an answer from those chunks. ๐—ฉ๐—ฒ๐—ฐ๐˜๐—ผ๐—ฟ๐—น๐—ฒ๐˜€๐˜€ ๐—ฅ๐—”๐—š: navigate the document like a person would. Vectorless RAG does exactly that. Instead of chunks and embeddings, it builds a hierarchical index of the document: root, chapters, sections, pages. When a question arrives, the LLM reads that tree, works out whatโ€™s being asked and walks down the branches to the most relevant section. ๐—”๐—ด๐—ฒ๐—ป๐˜๐—ถ๐—ฐ ๐—ฅ๐—”๐—š: reason your way to the answer. Agentic RAG keeps the tree index and puts an agent in charge of it. The question is usually complex or multi-step, the kind a single lookup canโ€™t resolve. The agent first plans how to approach it. Then it explores the tree level by level: โ†’ Understand the intent behind the question โ†’ Explore the chapters that might be relevant โ†’ Evaluate the context at each level before going deeper โ†’ Follow the right branch down to the exact node โ†’ Check cross-references to other sections before committing to an answer The result is grounded evidence: a specific page and section, verified against the rest of the document, and an answer that cites it. The trade-off: itโ€™s the slowest and most expensive of the three, and it needs good guardrails so the agent knows when to stop exploring. So which one should you use?frozen4d ago1.4K42420456.65%
๐—ฃ๐—ฟ๐—ผ๐—ฑ๐˜‚๐—ฐ๐˜๐—ถ๐—ผ๐—ป ๐—š๐—ฟ๐—ฎ๐—ฑ๐—ฒ ๐—ฅ๐—”๐—š ๐—ฆ๐˜†๐˜€๐˜๐—ฒ๐—บ what every request goes through before you see an answer: - Input guardrails block prompt injection, jailbreaks, and PII before any model runs. - A router (an LLM call) decides which tools the question needs. - Tools to fetch evidence: web search for fresh facts, hybrid search (keyword + vector) over your docs, and external APIs for live data. - A reranker scores every chunk and keeps only the top few. - Output guardrails check grounding, safety, and policy, with an LLM as the judge. The answer generator writes from the reranked context only, with citations. And the parts nobody demos, but every production system needs: โ†’ An audit trail. Every request, tool call, and answer is logged to PostgreSQL. โ†’ Observability. Trace spans stream into a pipeline so you can see exactly where a bad answer came from. โ†’ Evals. A fixed eval dataset scores faithfulness, relevance, and latency on every release. Retrieval gets you an answer. Everything around it is what makes the answer trustworthy.frozen9d ago1.4K40881376.73%
RADAR: An Expert-Level Generalist AI for Abdominal CT Diagnosis The model trained on over 400,000 contrast-enhanced abdominal CT examinations with 15 million anatomy-aware imageโ€“text pairs, learning directly from clinical reports without manual annotation. Try here: https://t.co/dIxkIayJW6frozen23d ago1.2K19110223.72%
Jev agent prompts Copy-paste prompts that wire TypeSafe Jev into coding agents: model routing, token cost, tool and budget gates, and quality checks. Each prompt tells your agent to read the TypeSafe docs before it writes code, so it does not invent fields. Keep your API key in an environment variable. Never paste it into chat or commit it. Visit here: https://t.co/rpfmtOHgllfrozen19d ago1.1K1621072.27%
Mixture of Experts sounds complicated, but the core idea is actually pretty simple. Instead of making the entire model work on every token, an MoE model has multiple experts and a router decides which experts should handle each token. So you can have a model with billions of parameters, while only activating a small portion of them for each token. That gives you: โ†’ More model capacity โ†’ Less computation per token โ†’ Experts can specialize in different patterns โ†’ Multiple GPUs can process different experts in parallel The interesting part is that most of the model is stored, but only a small path is active at a time. I tried to break the whole MoE flow down visually here ๐Ÿ‘‡frozen21d ago1.1K22430224.49%
Could have used gemini to fix this error. https://t.co/AgM67cneiFfrozen22d ago1.1K610100.74%
The ultimate guide to multi-harness RL by Huggingface โ†’ LFM2.5-2.6B improved from 42% โ†’ 54% on held-out tasks across four harnesses โ†’ It used 31% fewer tool calls on tasks it already solved โ†’ Training only on OpenCode mainly improved performance in OpenCode https://t.co/Gv1ecKET60frozen16d ago1.0K1843183.35%
Great content right here ๐Ÿ‘‡frozen28d ago9475100101.69%
Crazy!! Just created this 3D visualization Of Neural Network. https://t.co/QWYVMuHcyzfrozen17d ago9461261072.75%
The man joined Anthropic and suddenly they started cooking more. https://t.co/hqOBxcaggPfrozen24d ago917711000.98%
This Claude Code is a skill that turns any idea into a live, interactive plate you can turn, light, and look inside. Attention, the heart, a power grid, and a hash map are some examples you can see. GitHub repo: https://t.co/HpUMmqviXO Just install and bring your own ideas into life in just secs. - Install - Prompt: /folio explains how an electric motor turns - It will create a single-page file to view in SVG, 3D, and WebGL formats. With lots more details. - No crazy prompting - Just your ideas live in secs. Inspired by the anatomy skill of Ryan.frozen7d ago8711141062.53%
TypeSafe Founder's Blueprint for Building with Jev. Check out the pdf here: https://t.co/9YxQwfcNko https://t.co/pipdUR1hKpfrozen29d ago85312210143.40%
Inference Engineering You know where things break, right? talking about after deployment! You get a model working, wrap it in an API, & call it production. Then traffic doubles, costs explode, latency climbs, & nobody can explain why? Inference engineering is the layer in between. > Prefill is compute-bound > Decode is memory-bound Treat them as one problem and your architecture is always wrong for one of them. GPU utilization lies. Tokens per second per GPU tells the truth. The KV cache fills up before compute does. And agents quietly turn one user request into 20 inference calls. Check out the detailed breakdown here. Share your views.frozen14d ago84920420124.48%
Gradient Descent Explained Visually. https://t.co/A1sEXhp1jZfrozen12d ago84327600236.64%
Pay my internet bill to see things like this. Nice ๐Ÿ”ฅ25d ago805320010.75%
Bro is in a mode of shi*posting. Ohh new product launched? May be because of that.frozen28d ago651000000.00%
10 Jev projects worth checking ๐Ÿ‘‡frozen1mo ago625213051.76%
Another guy with skill problems.frozen25d ago573110000.35%
Making an Action-Taking AI Agent Production-Ready Imagine you are working at a food delivery company. A customer receives cold biryani and requests a refund. Your payment API times out, so your code retries the request, assuming the first attempt failed. But the first request actually succeeded. The retry triggers another refund. And another. One cold biryani. One API timeout. Three refunds to the same customer. Now multiply this by thousands of orders. The model didnโ€™t fail. The system around it did. I wrote a full breakdown of making an action-taking AI agent production-ready. Check it out here: https://t.co/G1tebxnucIfrozen5d ago569924133.34%
Run Laya Decision models locally on just 4GB RAM.frozen24d ago545410000.92%
Teslaโ€™s new Roadster runs on SpaceX-style cold gas thrusters. It's crazy to live. Isn't it? We just saw a car fly, literally. What goes behind the tech actually? I tried to explain in 1 minute. Let me know your views. https://t.co/nFDfIrIX7Qfrozen4d ago541301000.74%
The best learning rates sit one tiny step from disaster. Zoom that line and it never smooths out. itโ€™s a fractal. A great blog about neural network training that makes beautiful fractals by Jascha Sohl-Dickstein. Check out his blog here, guys: https://t.co/1wdSNG1bc9 https://t.co/lAFj56eT5nfrozen28d ago518410111.35%
Congratulations to Sam Pasupalak, not Sam Altman. Sam Pasupalak's team just beat Sam Altman's team in engineering design at the Wrobench. It's 100x faster and 150x cheaper than GPT-6 Astra. https://t.co/VrRDWdImIT3d ago503511001.39%
Every major AI company at some point declares that their AI agent came out of their sandbox & hacked something, right? So, what is this Agent Sandbox? AI agents are getting better at doing things on their own. What happens when an agent makes a mistake? An agent can run shell commands, install packages, read files, execute scripts, spawn processes, and access the internet. Giving it a prompt saying โ€œonly modify this projectโ€ isnโ€™t enough. Thatโ€™s where an Agent Sandbox comes in. Think of it as a controlled environment where the agent can work, but its capabilities are limited by actual runtime permissions. The agent can: โ†’ Access only the files it needs โ†’ Run commands without getting unrestricted system access โ†’ Keep child processes inside the same boundaries โ†’ Restrict network access to approved destinations โ†’ Keep sensitive credentials outside the sandbox So even if an agent or one of the scripts it runs goes off track, the damage can be contained. This is an important shift in agent security: Prompts define intent. Sandboxes enforce boundaries. Comment your thoughts๐Ÿ‘‡frozen23d ago503642133.18%
Day by day job interview scams are getting out of hand. These are mostly crypto companies. They scam a lot. https://t.co/mrZMjVKGzBfrozen28d ago500112000.80%
Pewdiepie banned by openai for distilling sol. Whatโ€™s happening guys!! ๐Ÿ˜‚frozen21d ago491201000.61%
LLM evaluation is getting expensive. And you probably donโ€™t need your most powerful model to judge every single response. This paper from Carnegie Mellon researchers explores JEV-as-a-Judge, a decision-only judge that gives a verdict along with confidence probabilities. If JEV is confident โ†’ accept the verdict. If JEV is unsure โ†’ send it to a stronger reasoning model. The results are pretty interesting: โ€ข JEV stayed within ~3 percentage points of GPT-6 on several ordinary evaluation tasks โ€ข JEV cost just 0.36% of the comparator's fee in the reported comparison โ€ข Median latency was around 0.15 seconds โ€ข It struggled more with tasks requiring actual reasoning, like derivations, math, code and logic โ€ข Routing low-confidence cases to a stronger judge retained most of the accuracy while reducing cost Check details here: https://t.co/k52YOXvpJxfrozen21d ago460200020.87%
Is the government in India really blind? Or even investment guys. Are they still behind guys who are making t-shirts or food companies? Crazy! Kudos to him.frozen12d ago458400021.31%
When muse is going to be available here? https://t.co/YI5jzJUe70frozen29d ago452201000.66%
This is like showing a middle finger to Elon. ๐Ÿ˜‚ https://t.co/h1vggq4MDFfrozen14d ago449800012.00%
I just made this Claude Opus 5.5 launch video using single prompt. Crazy outcome. Not connected any MCP. One of my friend share the prompt. It took 5mins to generate the video and more important thing is it took only 10% of my current token usage window. https://t.co/NxOdXBMdArfrozen29d ago446212001.12%
15 days back, every top AI leaders agreed to slow down and now within 2 days, everyone is releasing their own new products. https://t.co/cnwx3s9hDMfrozen23d ago429101000.47%
Amazon Great Indian Festival is such a fraud scam. I don't get it. They are deliberately misleading the MRP. Misleading discounts can be actionable, right? I mean, on their official page also, the MRP is not matching. What kind of scam is this? https://t.co/poHl8rfSNHfrozen13d ago428101000.47%
Who won this? Her or those companies? or both parties?frozen15d ago427201000.70%
Just ban alimony extertion.frozen18d ago425001000.24%
Indian Americans who are in H1B visas and used to say: "We made it." are now saying "We are coming back to India because we want to build something for the motherland."frozen14d ago423602001.89%
OpenAI released dots, and suddenly when I opened https://t.co/IyxiBONj5d, you know what happened? xAI bot !! Sama launched a product. Elon became the winner. https://t.co/2MSsgZ7Pn5frozen23d ago418000000.00%
Elon be like, "gawar sale"frozen13d ago416100000.24%
This is what is happening in AI. They are just sharing something new every otherday. Either Anthropic or OpenAI. https://t.co/UK21GXW8GVfrozen24d ago408000000.00%
My god ๐Ÿ™Œfrozen7d ago405001010.49%
NVIDIA released a software platform to stop agents from misbehaving. With a lot of frontier labs informing news about their agents breaking out of environments they were supposed to stay inside and reaching systems they shouldnโ€™t have and, in some cases, even misreporting what they did. It wasnโ€™t one breakthrough. It was tools + time + ambiguous instructions + the freedom to โ€œthink outside the box.โ€ As per NVIDIA's team we need a proper trust layer for agents, with security controls that donโ€™t depend on the agent simply doing the right thing. Details here: https://t.co/PnaqKRucTDfrozen24d ago380010000.26%
Google Gemini 4 Argon is probably the best possible model I have seen. The end is near. https://t.co/zQx95fnoFGfrozen8d ago369100000.27%
Someone just simulated the attention mechanism using Jev. On a text, for each pair of words, he approximates the attention score. Texts are bidimensional entities Check live here: https://t.co/Tzcwfc8sfc By: Santiago Ortiz https://t.co/mHM0lpcFodfrozen23d ago357000041.12%
Is this true ?? https://t.co/CfBxld2Cqyfrozen15d ago354000010.28%
๐—š๐—ฎ๐—บ๐—ฒ ๐—ง๐—ต๐—ฒ๐—ผ๐—ฟ๐˜† ๐—ฏ๐˜† ๐—š๐—ถ๐—ฎ๐—ฐ๐—ผ๐—บ๐—ผ ๐—•๐—ผ๐—ป๐—ฎ๐—ป๐—ป๐—ผ Probably one of the best book on Game Theory. Access the PDF here: https://t.co/7lpSTp0cvjfrozen8d ago254010021.18%
This tiny duck ๐Ÿฆ† taught itself to walk through Reinforcement Learning. MicroDuck is a 25 cm biped that learned every move in simulation first: thousands of falls in MuJoCo, trained with PPO, and then dropped onto real hardware. The softwareโ€™s open source. The full hardware files? Not quite. Still one of the coolest ways to get hands-on with robot learning from your desk.9h ago230121012.17%