Finteresting

5.4K posts

Finteresting banner
Finteresting

Finteresting

@FinterestingNow

#Fintech and financials... keeping it interesting with occasional shitposting

Katılım Nisan 2013
2.3K Takip Edilen193 Takipçiler
Finteresting retweetledi
TensorTonic
TensorTonic@TensorTonic·
You don't understand transformers until you've built one from scratch. Tokenization → Embedding → Positional Encoding → Scaled Dot-Product Attention → Multi-Head Attention → Feed-Forward Network → Layer Norm → Encoder → Decoder → Full Transformer. > Each block is a coding problem. > Each one runs against real test cases. > No IDE setup, no environment issues, just open and code. We broke "Attention Is All You Need" into subset of problems so you can build the entire architecture one piece at a time. Try it free: tensortonic.com
TensorTonic tweet media
English
6
90
1K
38.1K
Finteresting retweetledi
trace37
trace37@trace37_labs·
Fine-tuning Claude Code (or any LLM) to quasi-autonomously hunt bugs is (a) complex and (b) is primarily learned from agonisingly painful and bitter experience. Just one single skill (/sec-analyze) which takes js sinks etc and taint traces to user input is 728 lines... but it is the most successful skill I have for finding reportable bugs. This weeks swearing / frustration metric was 64 - down from 90 ish last week.. and down from 250 3-4 months ago.
trace37 tweet media
English
3
20
123
7.4K
Finteresting retweetledi
Baidu Inc.
Baidu Inc.@Baidu_Inc·
🚀 Introducing Qianfan-OCR: a 4B-parameter end-to-end model for document intelligence. One model. No pipeline. Table extraction, formula recognition, chart understanding, and key information extraction, all in a single pass. Paper: arxiv.org/abs/2603.13398 Models: huggingface.co/collections/ba… 🧵 Key results ↓
English
12
95
647
50.9K
Finteresting
Finteresting@FinterestingNow·
@brian_scanlan How do you protect personal info while giving production access to claude?
English
1
0
1
49
Brian Scanlan
Brian Scanlan@brian_scanlan·
The console is part of a broader Admin Tools MCP that gives Claude the same production visibility engineers have: Customer/feature flag/admin lookups etc. A skill-level gate blocks all these tools until Claude loads the safety reference docs first. No cowboy queries.
English
2
0
58
41.1K
Brian Scanlan
Brian Scanlan@brian_scanlan·
We've been building an internal Claude Code plugin system at Intercom with 13 plugins, 100+ skills, and hooks that turn Claude into a full-stack engineering platform. Lots done, more to do. Here's a thread of some highlights.
English
77
196
3K
783.6K
Finteresting retweetledi
ollama
ollama@ollama·
Nemotron 3 Nano 4B is now available to run via Ollama: ollama run nemotron-3-nano:4b Try it with Pi, the minimal agent runtime that powers OpenClaw: ollama launch pi --model nemotron-3-nano:4b This new addition to @nvidia's Nemotron family is a great fit for building and running agents on constrained hardware.
English
49
120
1.1K
62K
Finteresting
Finteresting@FinterestingNow·
This was one of the biggest ecological tragedy
Sama Hoole@SamaHoole

There were sixty million of them. That is not a round number invented for rhetorical effect. That is the estimate based on historical accounts, trading post records, early naturalist surveys, and the archaeological evidence of a grassland ecosystem that had been shaped, managed, and sustained by their presence for approximately ten thousand years. Sixty million bison, moving in herds so vast that 19th century travellers reported watching them pass for days without the column ending. The sound carried for miles. The ground vibrated. Early European explorers described riding to the top of a rise on the Great Plains and looking out at a sea of brown moving in all directions to the horizon, beyond which more was coming. These animals were not incidentally present on the Great Plains. They were the mechanism that made the Great Plains what they were. A bison herd moving across shortgrass prairie does something very specific. It grazes heavily, pulling the top of the grass. It aerates the compacted soil with hooves that break the surface crust and create small depressions, bison wallows, that collect rainwater and become micro-habitats for hundreds of species. It deposits dung that feeds a cascade of organisms from beetles to birds. It rolls in the disturbed soil, dispersing seeds in its coat across miles of subsequent travel. It moves on. This last part is crucial. The herd moves on. The grass it grazed comes back stronger. The roots, some of which extend twelve feet into the soil, deeper than the roots of any arable crop: draw carbon from the atmosphere and hold the topsoil together against drought and wind with a grip that the Great Plains had developed over millions of years of exactly this relationship. The Plains Indians who lived within this system understood it with the intimacy of people whose survival depended on it. They followed the herds. They took what they needed. They used every part of every animal: hide, bone, fat, organ, sinew, horn, dung, in a closed-loop material economy that generated essentially no waste. The calves born each year exceeded the animals taken by human hunters by a margin that kept the population stable at sixty million. This was a functioning ecological system that had been maintained in sustainable equilibrium for thousands of years. Then, in approximately thirty years, it was gone. The US Army did not accidentally allow this to happen. They planned it. General Philip Sheridan stated it openly: the hunters were doing more to "settle the vexed Indian question" than the entire military had managed in thirty years of direct combat. Every dead bison was a step toward starving the Plains nations into submission. Columbus Delano, Secretary of the Interior, articulated the logic without apology: "Every buffalo dead is an Indian gone." The hunters came. The railways came. Tourists shot bison from train windows. The carcasses were left to rot, stripped only of the hide and the tongue. Within thirty years, sixty million animals had been reduced to approximately three hundred. The Plains grasslands, stripped of the animal that had managed them for ten millennia, began to change. The deep-rooted perennial grasses that had anchored the soil slowly gave way to annual species less able to hold topsoil under drought conditions. Settlers ploughed what remained. Monoculture wheat replaced the native grassland complex. In the 1930s, the topsoil of the Great Plains blew away. The dust clouds reached Washington DC. The Army had solved the Indian question. It had also, by removing the ruminant that maintained the grassland, created the conditions for one of the worst agricultural collapses in American history. The sixty million bison were not causing the planet to overheat. The sixty million bison were the planet's solution to the problem we have since made considerably worse. They're doing their best to make the same mistake again with cattle.

English
0
0
0
3
Finteresting retweetledi
emergence.ai
emergence.ai@emergence_ai·
Today, we’re excited to introduce Emergence India Labs: India’s first dedicated frontier AI lab focused on autonomous agents, based in Bengaluru.
emergence.ai tweet media
English
8
19
149
40.1K
Finteresting retweetledi
Benjamin Marie
Benjamin Marie@bnjmn_marie·
Nemotron 3 Super GGUF evals: Q2 actually works. Model size goes from ~242 GB → ~53 GB, with only a tiny quality hit. Super impressive quantization robustness here, likely helped by NVFP4 pre-training, plus some very real Unsloth magic on top 🪄 Note: As usual, don't over interpret these results. +/- 10 points of relative error increase is not very meaningful. But above 10 points of increase, you can consider the model is bad.
Benjamin Marie tweet media
English
8
18
143
8.4K
Finteresting retweetledi
Tencent AI
Tencent AI@TencentAI_News·
For everyone who's been asking — Tencent #QClaw is opening up more spots in the beta 🦞 Built on OpenClaw, designed for everyone. Zero config, 3-step setup, runs locally. What's new: 🔹 WeChat Sync: Deeply integrated with WeChat Mini Program, enabling mobile command & seamless file transfers (Voice & image inputs coming soon). 🔹 Insights Hub: Relevance, auto-loaded. We've ditched the manual prompts. Relevant skills appear automatically based on your context. One tap to activate. 🔹 Local First: Secure, fast, and now available for download in China, with more regions to come. Join the Beta (plenty of invites!): qclaw.qq.com
Tencent AI tweet media
English
57
52
289
53K
Finteresting
Finteresting@FinterestingNow·
Interesting take on inferencing chips by Nvidia
Andrew Feldman@andrewdfeldman

NVIDIA's biggest GTC announcement was a $20 billion bet on the same problem we solved 6 years ago. Their next-gen inference chip - not available yet - has 140x less memory bandwidth than @cerebras. To run a single 2 trillion parameter model, you need 2,000+ Groq chips. On Cerebras, that's just over 20 wafers. Even paired with GPUs, Groq maxes out at ~1,000 tokens per second. We run at thousands of tokens per second today. And every day. In production now. Why? When you connect 2,000 chips together, every interconnect has latency. Every cable has overhead. It doesn't matter what your memory bandwidth is on paper if you're bottlenecked by the wiring between thousands of tiny chips. We solved this with wafer scale. One integrated system. Little interconnect tax. Jensen told the world that fast inference is where the value is. He’s right - it’s why the world’s leading AI companies and hyperscalers are choosing Cerebras.

English
0
0
0
14
Finteresting retweetledi
Anaconda
Anaconda@anacondainc·
Big news: Anaconda’s collaboration with @NVIDIA now spans the full enterprise AI stack—from GPU-accelerated Python environments to open models for agentic AI. Learn more about the expanded partnership: bit.ly/3Nmv1Qy
Anaconda tweet media
English
0
4
3
665
Finteresting retweetledi
Alex Prompter
Alex Prompter@alex_prompter·
🚨 BREAKING: NVIDIA sold the most powerful AI chip ever built. Then Princeton discovered the software running on it was wasting 60% of it. Every inference job. Every training run. 60 cents on every dollar, gone. > NVIDIA doubled the raw compute power of their Blackwell B200 GPUs compared to Hopper H100. Tensor core throughput went from 1 PFLOPS to 2.25 PFLOPS. The most powerful AI chip ever built. > The problem: the rest of the chip didn't scale with it. Memory bandwidth stayed the same. The exponential unit stayed the same. So the bottleneck moved and all that extra compute sat idle while the slower parts of the chip became the new ceiling. > Every existing attention implementation, including FlashAttention-3, was designed for Hopper. On Blackwell they either left massive performance on the table or couldn't run at all. > Princeton, Meta, and Together AI spent months redesigning attention from scratch around the new bottleneck. New pipelines. Software emulated exponential functions. A completely different backward pass. The result: FlashAttention 4. → Up to 2.7× faster than Triton on B200 GPUs → Up to 1.3× faster than NVIDIA's own cuDNN library → Reaches 1,613 TFLOPs/s 71% of theoretical maximum → Compile time dropped from 55 seconds to 2.5 seconds (22× faster) → Written entirely in Python no C++ template expertise required The scariest part: this wasn't a hardware problem. The chip was delivering exactly what NVIDIA promised. The software just wasn't designed for it. Every AI lab running B200s before this paper was paying for compute they couldn't use.
Alex Prompter tweet media
English
42
63
403
36.7K
Finteresting retweetledi
Alexander Doria
Alexander Doria@Dorialexander·
Breaking: @pleiasfr and @nvidia release the first open synthetic dataset for personas in Europe: Nemotron-Personas-France. 1M synthetic French persons, with rich imaginary lives grounded on (complex) demographic distribution.
Alexander Doria tweet media
English
33
90
714
76.8K
Finteresting retweetledi
Marktechpost AI Dev News ⚡
Mistral AI Releases Mistral Small 4: A 119B-Parameter MoE Model that Unifies Instruct, Reasoning, and Multimodal Workloads Key Differentiators: → Mistral Small 4: One model to do it all. → 128 experts, 119B total parameters, 256k context window → Configurable Reasoning → Apache 2.0 License → 40% faster, 3x more throughput Full analysis: marktechpost.com/2026/03/16/mis… Model on HF: huggingface.co/collections/mi… Technical details: mistral.ai/news/mistral-s… @MistralAI
Marktechpost AI Dev News ⚡ tweet media
English
2
7
18
604
Finteresting retweetledi
Xenova
Xenova@xenovacom·
IBM just released Granite 4.0 1B Speech, a compact and efficient speech-language model, designed for multilingual speech recognition and bidirectional speech translation. New #1 on the OpenASR leaderboard! It can even run in your browser on WebGPU, thanks to 🤗 Transformers.js
Xenova tweet media
English
5
53
255
20.5K
Finteresting retweetledi
ModelScope
ModelScope@ModelScope2022·
🔥 Meet dots.ocr-1.5: 3B OCR model from Rednote-hilab , SOTA multilingual document parsing, virtually any writing system. 📊 Elo 1089 on olmOCR-Bench, 1157 on XDocParse — above GLM-OCR, and PaddleOCR-VL-1.5 📄 OmniDocBench text edit 0.031, beats Qwen3-VL-235B (0.069) and Gemini 2.5 Pro (0.075) 🎨 SVG code output for charts, diagrams, and chemical formulas 🌐 Web parsing, scene text spotting, and object counting included ⚡ vLLM supported, single GPU 🤖 Model: modelscope.cn/models/rednote… 🔗 GitHub: github.com/rednote-hilab/… 🎠 Demo: dotsocr.xiaohongshu.com
ModelScope tweet mediaModelScope tweet mediaModelScope tweet mediaModelScope tweet media
English
1
36
244
11.3K