Lollipop

@above_spec Ah ok, thanks!!, CPU 9950x . 32 v threads but Is mem (ddr5 6000) band capped (~70gb/s), so sweet spot 8 -t 8 -tb 8 -c 0-7 -ub 32 -fa on --no-mmap -ctv q8_0 -ctk q8_0 -muge -no-fug. For the prefill -cram 8192 and --keep are required to have cache hits on system prmpt/tool desc.

@lollipop_stat Incredible! What CPU? So far I have only seen 3090s and higher do speculative decoding. My limited testing on 16gb gpus hasn't unlocked any gains. Maybe it would work on smaller 9b models. 0.8+9b

RTX 5060 Ti 16GB. $429 GPU. Last night I got 128 t/s on Qwen3.6-35B using ik_llama.cpp's R4 quant format. Crushing performance. Faster than the 5070 Ti on mainline llama.cpp. Performance stays consistent from 0 to 139k context and no speculative decoding used!🤯 Special thanks to @MakJoris for sharing ik_llama.cpp with us! Today I wanted to know if it's actually *useful* at that speed. So I gave it a coding agent and 4 creative challenges. Here's what it built. 🧵

@0xSero How are we sure chat sessions personal add value? It won t be about knowledge as the think we inference are part of the latent space. the only value maybe is training LLMs to Better orchestrate subagents and to reduce the retry loops, bugs created, etc.

Do you want to learn to use AI, and contributed your session data to open source so we can train better models? Models better than Opus We need as many people as possible to contribute their agent traces from their claude code + codex history Pi's Mario & I both shared ours.

@0xSero I got 50% k v savings, but quality degrades significantly (garbage output) from Q4 downwards. Till q5 models It has 1% error.

I need someone who's not been one-shotted to validate these numbers, I have gotten consistent results but I don't have enough expertise to make sure the models are honest. github.com/0xSero/turboqu…

@ossamandere @__tinygrad__ wow 100-200k context window is quite remarkable. i am about to do same build :), yes we need probably over 100tok/s to get it working

@lollipop_stat @__tinygrad__ A little too slow for live use, but if you hook it up in a ralph loop with a decently large context (100-200k tok) it tends to stay on the rails pretty well. Mighttake a while tho. Trying to get to 100 tok/s, that's where the response is really great for live

To anyone using local LLMs, what hardware do you have, what model are you using, and what tok/s are you getting?

@ossamandere @__tinygrad__ thank you! are you able to use it as 'agentic' or too slow to chain and orchestrate different LLM calls and their context?

@__tinygrad__ dual amd ai pro 9700s running qwen 27b q8 rocm backend getting 26 tok/s

@FinanceFreeman it is not the subsidy cost the problem, but price multiplier that is shrinking each cycle: The math that saved Bitcoin mining profitability before was: Subsidy cuts by 0.5x, but price goes up 5-10x = net positive, currently price up < 1.2, which is the reason of miners unprofits

🚨Bitcoin’s Catastrophic Problem NO ONE Is Talking About! Bitcoin produces 144 blocks per day. Fees are paid per block, every ~10 minutes. Yesterday, Bitcoin Layer-1 fees totaled ~2.6 BTC for the entire day. (Approx. $180K) That works out to ~0.018 $BTC per block. Meanwhile, the real cost to secure a block (electricity + hardware) is often $150k–$200k, depending on hash rate and energy prices. Today, miners survive because of the 3.125 $BTC block subsidy. Fees do not matter NOW, because the Bitcoin emissions are worth enough to offset costs. 🚨THE PROBLEM? When Bitcoin reaches 21 million supply, that subsidy goes to zero. At that point, fees must fully pay miners — or hash rate drops, security weakens, and the system degrades. To break even long-term, Bitcoin needs tens of millions per day in L1 fees. Yesterday was sub $200,000. I own Bitcoin. I want it to succeed. But pretending this isn’t a real issue won’t solve it.

@saxena_puru AMD

🚨A new Trading Alert is now live📈 We've just invested in an exciting business which we'd been tracking for months (stock has just pulled back 50%). Its revenue is projected to grow by ~25+% CAGR for 3 years. Subscribers: Login 👇alphatarget.com

@slackercapital what is the alternative to gold for that same historic allocation purpose?

@saxena_puru why do you create twice the same posts?

Google Cloud grew +48%YoY, huge acceleration! Whisper number was +38%YoY. We were told AI is a bubble and the cloud is dead.

@saxena_puru hello Puru, why do you post twice the posts in the last 2 weeks? e.g. x.com/saxena_puru/st… and x.com/saxena_puru/st…

Jensen Huang: Market is wrong about software stocks "The notion that AI is somehow going to replace software companies is the most illogical thing in the world and time will prove itself" Interview date: 3 February 2026

@slackercapital why specifically today?

@LouisKnightWebb Try Kimi thinking 2, Better thank sonnet 4.5 in my tests

Aaand I've switched back to Sonnet 4.5

@StockShark16 false alarms can still precede true signals

The Hindenburg Omen is resurfacing again, so I felt compelled to post and debunk the myth. $SPY with Breadth Indicator: Market breadth is improving — about 52% of S&P 500 stocks are now trading above their 50-day moving average. When the market is on the verge of a real crash as in Feb 2020, breadth usually deteriorates well before prices nose-dive. But that’s not what we’re seeing right now. This looks more like a technical pullback, not the start of a major breakdown. If breadth continues to firm up, the market could resume its upside as early as next week.

@grok @scotiacaper @ClaytonCharts @AndrewThrasher @grok why 3 days left for potential trigger? from what? and what is relation between the 5% canary and the hinbenburg omen cluster?

Hindenburg Omen #8 (2.5% Stockcharts signal) Note: we have 3 trading days left to trigger @AndrewThrasher 5% Canary signal on the S&P

@eve_silb @venturetwins the test has a high statistical significance and ~2% false positive rate (it is explained in the official page iirc) in my case i was 132 iirc :S

@eve_silb @venturetwins RAADS-R Score Interpretation Below 65: Likely not autistic (neurotypical range). 90–129: Moderate to strong likelihood of autism. 130+: Very strong likelihood of autism should prompt formal diagnostic assessment if autism is suspected.

Every autistic person I know loves AI. Many of them use LLMs more than anyone else I've seen - and find it incredibly useful in their personal and professional lives. I think it's weird we don't talk about this more.

@eve_silb @venturetwins to get an idea if you could be autistic, try the RAADS-R test, ASRS and AQ they are free and few questions, let LLM help you understand the results

@venturetwins my psychiatrist thinks i'm probably autistic, but i'm not sure sure

Here Comes Another Bubble (AI Edition)

@grok @coinsavage @BackStCrawler @SubuTrade And what Is historic probability of correction After a cluster of 3 or 5 signals within the 30d period?

Definitions vary, but many sources define a Hindenburg Omen cluster as 2 or more signals within 30-36 days, with 3+ often seen as stronger (e.g., SentimenTrader cites 5 as notably ominous). In this thread's context, the 3 signals suggest a forming cluster. Always verify with current data.

A 3rd Hindenburg Omen triggered today, after 2 were triggered last week. We're starting to see a cluster of signals... 🚨