forget the grind. small iterative steps. do things (@halftroll) header
forget the grind. small iterative steps. do things (@halftroll) profile picture

Programmer/builder seeing world through lens of climate emergency, yet thrilled by the positive aspects of AI and general purpose technology @Blaine@sfba.social

SF Bay, China, Mexico, Italy
Joined September 2008
Photo by forget the grind. small iterative steps. do things (@halftroll)Photo by forget the grind. small iterative steps. do things (@halftroll)Photo by forget the grind. small iterative steps. do things (@halftroll)Photo by forget the grind. small iterative steps. do things (@halftroll)Photo by forget the grind. small iterative steps. do things (@halftroll)Photo by forget the grind. small iterative steps. do things (@halftroll)Photo by forget the grind. small iterative steps. do things (@halftroll)Photo by forget the grind. small iterative steps. do things (@halftroll)
Replying to @halftroll
countless lives could have been saved. Ignoring women in STEM/STEAM has consequences. Her contribution went well beyond topics like fresh air and ventilation. They included data collection—another area where WHO, CDC, and medical industry have been weak.
5
1
54
Outrageous
这么优秀的一个孩子,也被 ICE 拘留了,实在难以理解。 Eva Yi Xie,14 岁从中国来到美国时,连英语都还说不好。四年后,她入围美国高中生顶尖科学竞赛 Regeneron STS 全国 40 强,并进入 MIT。2024 年以数学、计算机双专业毕业后,又到普林斯顿攻读神经科学博士。 这次,她刚在华盛顿参加完神经科学年会、展示了自己的研究,就在返程途中被 ICE 拘留。原本该回实验室继续做研究的人,如今却被关进了拘留中心。普林斯顿已为她安排律师, 全力营救。
2
forget the grind. small iterative steps. do things retweeted
Replying to @tlanderso
I would like as many people to have as many rtx pro 6ks as possible! I would even argue that ive gained more utility the more people that have these cards and just local AI in general. The community has made local AI faster and more effective.
1
15
760
forget the grind. small iterative steps. do things retweeted
There is literally no situation in which exploding 10,000 lbs of lead is a good idea. We need full transparency around how this decision was made. nytimes.com/2026/11/19/clima…
493
2,681
85
13,609
152,572
forget the grind. small iterative steps. do things retweeted
Looped models are weird. A layer can write to itself or read from "future" layers. @thinkymachines gave us early access to Inkling 2 Mini so we could study its inner mechanisms. We uncover how it coordinates information flow across layers, blocks, and loops. Studying Inkling 2 Mini gave us cause for optimism: as transformers continue to evolve, we believe our ability to understand their internal mechanisms can evolve with them. Read the full post: goodfire.com/research/interp…
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index. thinkingmachines.ai/news/ink…
13
145
15
1,112
74,826
I wish you were wrong.
"I wholly expect Rs and Ds to coalesce the AI "regulation" movement around sinophobia in the next few months, redirecting US anxiety around AI job loss to the "Chinese-dominated, un-American, AI industry" - with Anthropic stoking the sinophobia to hobble their competitors" hmm...
1
4
forget the grind. small iterative steps. do things retweeted
🚨 My classmates girlfriend just got detained by ICE, please please reshare. Attaching his message with her story and actionable numbers to call. She was detained at Washington airport after flying back from a conference, and her last message was a call for help through a Slack channel before losing contact. Eva is currently a PhD student studying neuroscience at Princeton.
20
41
2
251
14,288
forget the grind. small iterative steps. do things retweeted
I just saw the paper as a result of @martin_casado's tweet (didn't see the paper when it was originally posted in Feb). Indeed, the analysis is great and the figures really make the paper shine! Nice work, Dhruva Karkada, Daniel J. Korchinski, Andres Nava, Matthieu Wyart, and @yasamanbb! I love analyses that reveal even more about the amazing hidden structures in word embeddings. arxiv.org/pdf/2602.15029
Thank you for the high praise 🙏🏻
13
27
2
264
17,578
Local
We worked on optimising the hell out of Inkling-2-small Sonnet level performance on a laptop
2
forget the grind. small iterative steps. do things retweeted
American open-source AI isn't dead yet 😝😝😝 @thinkymachines is releasing Inkling 2 Mini as apache 2.0 today and getting to the pareto frontier for open agent models. Model: huggingface.co/thinkingmachi…
64
122
15
1,607
43,967
forget the grind. small iterative steps. do things retweeted
The real threat Zohran Mamdani poses is to the career politicians who’ve spent decades making promises, then making excuses, then telling people to vote harder next time, is that elected officials can actually do things that they promise, if they want to.
We're closing Rikers once and for all.
64
1,301
20
6,815
52,209
Nice explanation.
Replying to @b1rkh0ff
They aren't doing it for charity. DeepSeek, Qwen and Kimi still charge for APIs (DeepSeek and Moonshot both hit ~$1B ARR). Alibaba uses free Qwen weights to drive its AI cloud business, which is growing 50%+. Self-hosting the biggest models remains costly and complex for most users, so hosted inference stays the easy path. Open weights also buy global adoption, talent, feedback loops and a workaround for chip export controls. Some licenses now restrict large commercial hosts. It's ecosystem + services revenue, not pure giveaway.
5
forget the grind. small iterative steps. do things retweeted
This performance of this song is so perfect that no other versions can compare! Both girls are amazingly good! Everything is in the style of a bygone era: the dresses, the shoes, the hair, the makeup, the movements, the gestures, the voice, and even the timbre. Everything is just perfect!
27
238
15
1,153
48,433
I like testing a European model.
🇫🇷 A French AI company released a 7B model that runs at 72 tps on an 🍎 Apple MacBook Air with only 16GB RAM! - No discrete GPU needed. The 6-bit quantization is actually faster than the smaller 5-bit version! 🤷‍♂️ Olala-v1.0-mini-7A1B is the new model from OVHcloud 🇫🇷. It's an open-weight reasoning model built using a hybrid Mamba-3, attention, and Mixture-of-Experts architecture. And it was designed to run efficiently on local hardware. 👏 Model specs ... 🧠 6.8B total parameters ⚡ ~1B active per token 🏗️ 29 Mamba-3 + 7 attention layers 📚 32K context 🐝 Native tool calling 💭 Four reasoning-effort levels 🌍 Trained on 12.8 TRILLION tokens 🔓 Apache 2.0 Performace of OVHcloud's reported results on an M5 MacBook Air with just 16GB unified memory. GGUF → Size → Generation speed 🟢 Q8_0 → 6.9GB → 60 tps 🔥 Q6_K → 6.3GB → 72 tps 👈 🔵 Q5_K_M → 6.0GB → 44 tps And the Q6 version also delivers 932 tps of PP. These measurements used a 512-token prompt and 128 generated tokens. OVHcloud reports that all three quantizations perform within benchmark noise of the original BF16 model on its tested coding and mathematics evaluations. That means the 6-bit model provides the best measured generation speed wo a quality penalty. The model supports reasoning, multi-turn tool use, and agentic workflows. So locamaxxer details ... 💾 6.3GB Q6 GGUF 🍎 Runs on a 16GB Mac 🟢 NVIDIA GPU support 💻 CPU inference supported 🐝 Tool calling through an OpenAI-compatible API 🔓 Downloadable weights and runtime code ⚠️ The GGUF currently requires OVHcloud's custom llama.cpp fork. Standard llama.cpp, Ollama and LM Studio don't support this architecture yet. There's also an MLX implementation for Apple Silicon. Links in ALT
4
forget the grind. small iterative steps. do things retweeted
Calling a genocide a genocide isn’t just about accountability, it’s about prevention. Choosing not to call a genocide a genocide isn’t just an act of cowardice, it’s an act of permission. nytimes.com/2026/11/19/world…
464
2,212
68
11,518
117,572
Claudia is number one.
What this does is set the standard for everyone else in government: if she flies like this, you can’t possibly ask to fly business. As an aside: the president that bought that brand new airplane used to travel around the world, bringing family, friends and telenovela stars for free tourist trips.
1
forget the grind. small iterative steps. do things retweeted
⬇️
22
49
4
200
4,300
U.S. arrogance is often amazing. And depressing.
Huh…First the NYT’s “mysterious dams,” then CNN’s “dirty bomb,” and now Reuters with its cartoonish, Persepolis-style stereotypes. As a girl who lived in Iran for 18 years and has now spent almost 17 years in America, I’m sick of Americans pretending they have some moral superiority over Iranians. Your country has produced some of the most horrifying cases of serial killers, child kidnappings, prolonged captivity, torture, and sexual abuse I’ve ever heard of. Children abducted and held captive for decades, subjected to unimaginable abuse. The sheer brutality of these crimes is beyond comprehension. I don’t even need to talk about your p*do rings cause you know more than me. You have child marriage, honor killings, sexual exploitation, mass shootings, and deeply rooted sexism. I never faced discrimination in my education in Iran, yet I encountered it in America many times. Your surgical specialties are still boys’ clubs. Yet you have the audacity to lecture Iranian women about civilization and morality? The hell with the Islamic Republic. I’m exposing the hypocrisy of a media establishment that loves portraying Iranian society as backward while conveniently ignoring the horrors within its own borders.
1
10
Worth investigating.
🇺🇸 A new American open-weight AI model matched Qwen3.8-27B on the Artificial Analysis Intelligence Index. It beats Qwen3.80-27B in nearly all shared benchmarks (check out the deltas) And you can run it locally with just 16GB of RAM. 😳 @thinkymachines released Inkling 2 Mini, a 26.4B-parameter MoE model using a fascinating new looped-transformer architecture. Unsloth has already released GGUFs (see thread) Look at the benchmarks comparing it with Qwen3.8-27B ... Artificial Analysis Intelligence Index v5 🟢 Inkling 2 Mini → 20.1 🔵 Qwen3.8-27B → 20.1 same DeepSWE v1.1 🟢 Inkling 2 Mini → 48.2% 👈🔥 🔵 Qwen3.8-27B → 42.2% SWE-bench Pro v2 Hard 🟢 Inkling 2 Mini → 77.9% 👈🔥 🔵 Qwen3.8-27B → 73.8% Terminal-Bench 2.1 🟢 Inkling 2 Mini → 71.6% 🔵 Qwen3.8-27B → 73.0% The AA Index results are independently evaluated and the coding results are Thinking Machines' published benchmarks. It has only 24 transformer layers. 👈🔥 Instead of adding more layers and parameters, Thinking Machines loops the middle 12 layers 3x for every token. That's 48 layer executions using just 24 stored layers. Essentially, more internal computation without needing more unique model weights. And it's a MoE with just 5.4B effective weight activations per token, roughly 5× fewer than dense Qwen3.8-27B. GGUF ladder ... 💾 Q2_K_XL → 9.14GB 💾 Q3_K_XL → 13.63GB 💾 Q4_K_M → 16.61GB 💾 Q4_K_XL → 17.63GB Local AI requirements ... ✅ Runs on 16GB systems with smaller quants ✅ 32GB+ Macs supported via Ollama ✅ Text + image inputs ✅ Up to 1M-token context ✅ Apache 2.0 ✅ GGUF for llama.cpp ✅ Open weights for local inference and fine-tuning The full 1M context isn't realistic on a 16GB setup, and llama.cpp currently requires a build with looped-layer support. Thinking Machines even demonstrated the model building a playable flight simulator entirely offline on a MacBook Pro. 👈🔥
4
Be like Minnesota.
Democrats won big. Now it’s time to show that we can deliver in Washington. With a one seat majority in Minnesota we delivered: Free school meals Paid family leave The Child Tax Credit Free college for low income families Reproductive freedom Clean energy by 2040
4
forget the grind. small iterative steps. do things retweeted
She arrived in America struggling with English and became an MIT-trained neuroscientist at Princeton. Now ICE has her LOCKED in a Virginia detention center. Eva Yi Xie, a third-year Ph.D. student at the Princeton Neuroscience Institute, presented a poster at the Society for Neuroscience annual meeting in Washington on Wednesday. On Thursday, November 19, she was detained at Reagan National Airport, according to her lab. She sent her lab mates a call for help on Slack. Then they lost contact with her for hours. She is believed to be held at an ICE facility in Chantilly, Virginia. ICE's online detainee locator lists her as in custody but doesn't say where. ICE has not answered multiple requests for comment and has given no public reason for detaining her. Her immigration status has not been made public. Xie came to the U.S. from Shanghai in 2016 and at first struggled with English. At the Bronx High School of Science she became a finalist in the Regeneron Science Talent Search for genomics research aimed at improving heart disease treatment. She graduated from MIT in 2024 with a double major in computer science and math. She has interned at Meta, Microsoft AI and IBM Research, and she now studies how the brain makes decisions. She is also helping organize an international computational neuroscience conference in Edinburgh for 2027. "Here, if you are brave enough to step out of your comfort zone, you can definitely achieve something," she told her high school paper in 2020. Her adviser, Princeton neuroscience professor Carlos Brody, called her "a brilliant law-abiding student and wonderful person." "Detaining her, with the intent of deporting her, makes no sense whatsoever. It is cruel, and it is our country shooting itself in the foot," Brody said. She is the first Princeton student known to have been detained by ICE. Her detention comes in a year when ICE made more than 491,000 arrests, the most in the agency's history. Princeton has arranged lawyers for her, and they are working to get her released, as Luke Grippo first reported in The Daily Princetonian. Her lab is asking people to call their senators and urge them to demand her immediate release. #ICE
347
433
41
1,728
177,967