logicgate retweeted
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…
Thinking Machines is the one of the few American organisations with a moral backbone
It is our duty to help protect and support them and spread their work!
Inkling mini matches Opus 5.5 on Cyber bench
logicgate retweeted
Update on my Google AI Ultra subscription nightmare.
My paid Google AI Ultra subscription was overridden by a free one-year promotional subscription, and things have gotten even worse.
Now my Google Colab subscription has also been canceled, and I can no longer use Colab at all.
The free subscription cannot be terminated early, no matter what I try. Even worse, the web-based subscription has taken control over my existing iOS subscription. I can no longer change my plan through Apple. The only option available is to cancel it.
I've contacted multiple Google support agents, and every single one has given me the same dismissive, scripted responses. None of them have actually addressed the underlying bug.
They claim they cannot manually remove the free subscription, cannot bypass the system, and apparently cannot even escalate the issue to a supervisor, engineering team, or anyone capable of fixing it.
Seriously, @GooglePlay @GoogleAI @GoogleAIStudio @googledevs , is this the level of customer support you provide to paying Ultra subscribers?
I'm literally paying for a premium subscription, yet a FREE promotion has stripped away my paid benefits, canceled my Colab access, and left me with no way to fix my account.
And your support team's response is essentially: "There's nothing we can do."
How is this acceptable?
Replying to @Google
@Google @GeminiApp This is unacceptable. I'm paying for Google AI Ultra, yet activating a FREE trial caused me to LOSE access to features I've already PAID for.
Here's what happened, and why I strongly recommend other Google AI subscribers avoid making the same mistake.
I already had an active Google AI Ultra monthly subscription through Apple's App Store. Everything worked perfectly, including access to Claude Opus 5.5 and Sonnet 5.5 in Google Antigravity.
Then I accidentally activated a promotional free trial of Google AI Ultra through Google Play.
And somehow, Google's subscription system appears to have prioritized the FREE trial over my existing PAID subscription.
The result? Antigravity no longer recognizes me as an eligible paid subscriber for third-party models.
I'm literally paying Google for AI Ultra while being treated as if I'm not a paying customer.
But the worst part is what happened when I contacted Google One Support.
After explaining the issue repeatedly, I was told:
They cannot manually remove or revoke the free trial.
They cannot restore my existing paid subscription benefits.
I should simply wait until December 17, 2026, for the canceled trial to expire.
Alternatively, I could DELETE my entire Google One service, potentially losing my existing membership benefits and account-related data.
There is apparently no direct channel for users to report this issue to the appropriate technical team.
Seriously, @Google?
Your system allows a free promotional subscription to override an existing paid subscription, and your solution is to tell the customer to WAIT or DELETE their Google One service?
Why should paying customers bear the consequences of what appears to be a bug in Google's own subscription entitlement system?
I don't want to cancel Google AI Ultra. I actually rely on Antigravity and its AI models for my work. That's precisely why I'm paying for the subscription in the first place.
I'm not asking for free services, special treatment, or additional benefits.
I JUST WANT ACCESS TO THE FEATURES I'M ALREADY PAYING FOR.
And to everyone using Google AI Pro or Ultra:
⚠️ PLEASE BE CAREFUL WITH GOOGLE AI FREE TRIAL OFFERS.
If you already have a paid subscription, activating another promotional trial on the same account may cause unexpected subscription conflicts.
Canceling the trial may only disable automatic renewal, meaning the conflicting trial benefits could remain active until the expiration date.
You could end up in the same situation as me: paying for AI Ultra while losing access to some of its paid features, with no immediate solution from support.
I've documented the entire issue on Google's official AI Developers Forum:
discuss.ai.google.dev/t/bug-…
@Google @GeminiApp Please investigate this subscription entitlement issue, provide an actual escalation path, and stop leaving paying customers without the services they've purchased.
A free trial should NEVER make an existing paid subscription less useful.
This needs to be fixed.
logicgate retweeted
We're hiring for an AI Platform Engineer on my team.
You'll work with me and my team directly enabling the org and implementing AI solutions and much much more all across the org.
100% remote. Details below.
Apply here: jobs.ashbyhq.com/polygon-lab…
logicgate retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me.
I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA.
I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation.
Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.
Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours.
Jensen
logicgate retweeted
Kicking off @foresightinst’s Vision Weekend in SF surrounded by frontier researchers and scientists. Some are working on reversing aging, some mapping the whole planet, some building brain implants or trying to figure out if AI can be conscious.
Our team at @eigenlabs is here because we believe that this kind of research should be done in the OPEN. Also the core idea behind @YukonResearch.
Fittingly, we’re at the Internet Archive who have archived over a trillion webpages!
Big thank you to @allisondman and the Foresight team for pulling this together.
logicgate retweeted
Today's Jensen Update Thread
1. WSJ fabricated statements from the Singapore government
logicgate retweeted
Replying to @WSJ @WSJopinion
You sort of set him up, didn’t you, by creating the impression that senior government officials anonymously provided information about spying, when in fact the info came from a lobbyist at a competing AI company?
logicgate retweeted
To my friends and family (and social media verse). I am getting slammed with conspiracy theories out of the park right now.
Everything is fine, take a breath - the systems will be operational soon and it'll be fine in the next few days.
You don't need to buy all the toilet paper out at the moment.
logicgate retweeted
Local AI Builders Friday # 4
The list is now 122! A lot of you nominated people under # 3, thank you ❤️
I added 11 more amazing builders this week:
1) @dhh is an OG! He is working to transform Omarchy into a Local AI distro. Starting with Omarchy 4.6 - it’ll try to select the optimal AI model for your hardware and lets you share AI models across your tailnet (ie use your DGX Spark to provide AI capabilities to your netbook)
2) @0xSero (@exolabs) got DeepSeek running entirely offline on his pack-of-portable-compute (that he literally carries around in a backpack!) 🤯 and is getting 200+ tok/s
3) @ItsmeAjayKV is doing his own Local AI R&D (kernels, quants, etc.) using just a 3090 and 64GB of RAM - and is hands-down one of the best open weight model news aggregators/accounts to follow
4) @christopherowen runs TP4 on 4 DGX Sparks cabled in a ring (no switch!) and implemented his own one-shot RoCE all-reduce algorithm getting 13-19 µs allreduce latencies (vs 75-86 µs for NCCL). Check out his repo with 279 experiments (he also logs his misses, with numbers!)
5) @tpae is building @OsaurusAI, free, open source (MIT) software for running local AI on your devices. Check out his Apple @ Work appearance this week (his pitch: “own your AI. At work too.”)
6) @dealignai (also working on Osaurus AI!) makes the case that since real tasks are input-heavy, using prompt processing on 4 DGX Sparks instead of a 512GB M5 Ultra makes sense (similar $)
7) @tayarndt and 8) @mikedoise are building Perspective Intelligence at Techopolis - an on-device AI chat assistant for Mac, iPad and iPhone. @tayarndt is also behind DictateMute - free, open source software that automatically mutes you on Discord when you’re dictating to Claude Code
9) @thelogicalgate uses his 128GB M5 Max MBP and 128GB RTX Spark to run local LLMs and is working on his own inference engine for RTX chips (to run on Windows on Arm)
10) @aijoey is running some of the frontier open models on just 2 DGX Sparks and documenting everything (configs, benchmarks, even building tools like Spark Doctor to help others get the most out of their Sparks)
11) @rickmanelius has 3 Macs + 2 DGX Sparks all networked together using Tailscale so he always has access to 3 or 4 local AI models (and can decide which one to use for different tasks)
Follow them and follow the full list here: nitter.gq/i/lists/21089799789857…
Who should be in # 5?
Local AI Builders Friday # 3
The list is now 111. I learn more from this list than from the rest of my timeline these days!
I added 7 more amazing builders to follow this week:
1) @niklaslenz_ai runs local AI at home on 4 DGX Sparks and an RTX 5090 and ranks what runs best on the machine you own. His test of Idol turned one image into a 3D model in 14.5 s on a 5090 and 57.4 s on a 16GB 5060 Ti. Now he's building a strategy game where local models fight each other on any hardware.
2) @plotarmordev is setting up a Mac and DGX Spark cluster at home and sharing what he learns. On one Spark, TensorFold took Qwen3.8 Flash Next from 75 to 95 tok/s vs vLLM at 2 users, with 34% faster prefill.
3) @brexton co-heads AI and compute infra at Cantor. In his view every founder he talked to in SF is working on owning their own hardware.
4) @javierbits builds local AI and security tools in public. He's testing LFM3-1.7B, and this morning he hosted an X Space teaching how to get started with local AI on whatever hardware you have. I caught part of it and it was great.
5) @summerAI_O is an 18 year old student (happy birthday!) in Japan training models on local + cloud. He posts daily breakdowns of new AI papers with the numbers and the caveats and yesterday he shipped Diffusion-Qwen3.8-27B, which fills in many tokens per pass instead of writing one at a time.
6) @NotASithLord is CTO of Hydra Host, which runs GPU clusters for people who want to own their own. His take is that context is the most valuable thing an agent has and memory is the bottleneck.
7) @Prajhan_026 trained HobbyLM-1B from scratch, a sparse MoE with ~305M active params, in 76 hours on 4 H200s. Now it does computer use fully offline at 0.26 s per decision, 13x faster than before. Deep into research.
Follow them and follow the full list here: nitter.gq/i/lists/21089799789857…
Who should be in #4?
logicgate retweeted
🚨CRITICAL MISSING PERSON🚨
NAME: America Marshall
AGE: 11
RACE/GENDER: Black Female
HEIGHT: 4’7”
WEIGHT: 80 LBS
EYES: Brown
HAIR: Black with green highlights, currently styled in a small afro
LAST SEEN:
2800 block of Wallace Covington Circle, Birmingham, Alabama
logicgate retweeted
The most ignored chip in your Mac is the Neural Engine.
LocalSR runs its upscaling models on it: 19 ms per tile on an M1 Pro versus 38 ms on the GPU, at lower power. Open source.
If a small model runs all day, it belongs on the ANE and the GPU stays free for your LLM.
nitter.gq/Hermes_Gy/status/21191…
On Apple Silicon, LocalSR runs its main upscaling models on the Neural Engine. On my M1 Pro a tile takes 19 ms instead of 38 ms on the GPU, using less power.
Free and open source: github.com/HerRei/local-upsc…
logicgate retweeted
Parallel drafters waste most of their guesses because token 9 has no idea what token 8 picked.
Tencent's DSpine passes each neighbor's guess forward through all 5 draft layers. On Qwen3.8-27B it accepts 7.20 tokens per round on GSM8K vs 5.69 for DFlash2, and agent replays run 2.7x to 3.8x faster than plain decoding on GPU.
Apache 2.0: huggingface.co/tencent/Qwen3…
logicgate retweeted
Stop converting GGUFs to MLX on your Mac.
MLX-Serve 26.11.2 runs DeepSeek-V4.1 GGUFs on its MLX engine (--mlx-gguf): ~3,300 tok/s prefill on an M5 Ultra, 2.5x the old path.
Same release: editing an old message in a 60k-token DeepSeek chat went from 8.7s to 0.13s.
logicgate retweeted
Coming soon, more ways to run Deepseek sub 128GB Ram on MLX-Serve.
Honestly this model is so good, at 96 - 128GB, It should be everyone’s new default.
Upcoming pruned iQ-MLX DS4 pack will run on a 96GB M5U with ioGPU set to 90, it will be tight, but fully in memory.
And... hoping for a Qwen's release soon, and it will make this post obsolete 🤞
logicgate retweeted
Everyone's best argument for 4x Spark over a Mac was prefill. oMLX just paired an M5 Max with an M3 Ultra and took DeepSeek V4.1 Flash from 1,060 to 3,123 tok/s at 64K. No new silicon. Told you software catches up. nitter.gq/jundotkim/status/21179…
oMLX 0.7.4.dev1 is out!
github.com/jundot/omlx/relea…
With the new Prefill Accelerator, connect an M5 Max to your big Mac and prefill up to 2.9x faster!
Prefill Accelerator (64K prompt, M3 Ultra 512GB + M5 Max 128GB)
- DeepSeek V4.1 Flash: 1,060 -> 3,123 tok/s (+195%)
- Qwen3.8 Flash-Next: 1,309 -> 3,666 tok/s (+180%)
- GLM-5.3 Flash: 540 -> 1,388 tok/s (+157%)
- MiniMax M3.1: 297 -> 569 tok/s (+92%)
- An M5 Max 64GB still adds 30-118%, and decode speed stays the same.
Faster DeepSeek V4.1 with A8 and buffer pool improvements
- Prefill on M5 Ultra is 43-48% faster, 32K goes 2,631 -> 3,779 tok/s (+44%).
- oQ4e builds now run INT8 Activation Prefill on M5.
Faster Qwen3.8 Flash-Next
- FP16 builds get the fused kernels. Decode on M3 Ultra goes 38.9 -> 63.0 tok/s (+62%), and +23% on M2 Max.
- Decode after long prompts is 3-5% faster on M5 Max.
Better for agents
- Parallel tool results from Claude Code reach the model in call order, so Qwen no longer pairs a result with the wrong call.
- Agent sessions on a cluster reuse the prompt cache across tool turns, 743 -> 9,266 cached tokens by the fifth turn.
- Switch oMLX models inside Codex and Copilot with /model.
Also in 0.7.4.dev1
- Ming-3.0 TTS with a 2x faster decode loop.
- Ask AI can use an oMLX server on another Mac.
- Check for model updates in the macOS app.
- German across the web dashboard and the macOS app.
- Batched decode with 6 requests on 8-bit models is 16% faster.
A big thank you to everyone who sent PRs, reported bugs and tested builds for this release, and a warm welcome to the five new contributors!

































