Every new comic strip is a small parallel reality.
Seattle
Joined June 2022
- Tweets2.1K
- Following199
- Followers22
- Likes8K
Gizmo Wizard retweeted
Ant open-sourced a 0.5B model that can clone a voice, design a new one from text, and make it laugh, sigh, cough, or whisper.
#1 open-weight model on Artificial Analysis Controlled Voice at 1069 Elo
The smallest in the family has these stats
🎙️ Ming-3.0-TTS-0.5B
→ 0.5B dense
→ zero-shot voice cloning
→ design a voice with natural-language instructions
→ control emotion, pace, pitch + emphasis
→ laughter, sighs, breaths, coughs, hesitation
→ 7 languages
→ 44.1 kHz speech
They released more than TTS.
🗣️ Ming-3.0-TTS-Tiny
→ 7.9B total / 1.3B active
→ multi-speaker dialogue + long-form speech
👂 Ming-3.0-Audio-Tiny
→ 9.3B / 2.7B active
→ understands speech, music + environmental sounds
→ spoken Q&A + dialogue
📝 Ming-3.0-ASR-Tiny
→ 8.6B / ~2B active
→ dialects, accents + contextual hotwords
And there's a much larger:
🧠 Ming-3.0-Audio-Flash
→ 126B / 6.5B active
🎯 So the entire stack can 🎤 hear locally ↓ 🧠 understand locally ↓ 🤖 agent decides ↓ 🗣️ speak back in a cloned/designed voice
No GGUF/llama.cpp yet, and no consumer-GPU benchmarks.
Links in ALT
ALT https://github.com/inclusionAI/Ming-3.0-tts https://huggingface.co/inclusionAI/Ming-3.0-tts-0.5B
Gizmo Wizard retweeted
#KempnerInstitue researchers consider how market-like rules can help groups of AI agents solve hard problems together... check out the animation below and our recent article about the preprint: bit.ly/4wOAwrI
Leonard Read's pencil story: no one knows how to make a pencil from scratch, yet millions are made without a central manager. Each worker knows only their part. Incentives, not central control, coordinate their efforts.
Current multi-agent AI seems to work the opposite way. A central manager does everything: assigns roles, writes the workflow, and decides who acts next. As agents increase and tasks get harder, that manager might become the bottleneck: it should understand every agent and every task, and adaptation becomes inefficient.
So how can a group of machine agents organize themselves into something none of them could achieve alone? Further, how can that society of agents keep improving itself over time?
In Economy of Minds (EoM), we show that simple local incentives can drive global self-coordination and self-improvement for modern LLM agents. We let agents live in a minimal economy: they compete for work, earn by helping others succeed, and go broke when they don't. Workflows, specialists, and improvement naturally emerge, with no one in charge.
Across 5 domains (math reasoning, financial research, scientific research, accelerator design, and distributed-system optimization), limited agents grow into strong societies.
The whole idea in a 90-second cartoon👇
Gizmo Wizard retweeted
Our research team did a study and found that optimizing your harness can drastically improve model performance to a shocking extent.
We ran frozen @NVIDIAAI Nemotron 3.5 Super VL weights through 9 off-the-shelf harnesses on the same 100 SWE-bench Verified tasks.
The results ranged from 30 to 76 tasks solved, a huge spread.
Each run used the same endpoint, 100 tasks, one attempt per task, and a 4,500-second limit, so the 46-task spread came from the harness.
We found similar results with other models:
- Claude Opus 4.5 went from 33.5% to 77.5% on BFCL V4 just by adding native tool schemas.
- A new execution pipeline moved GPT-4o from 23.2% to 38.8% on SWE-bench Verified (but used much more inference).
- AHE gained from tools + middleware + long-term memory on Terminal-Bench 2. Its system prompt alone hurt performance by 2.3 points.
If you’re evaluating an agent, it’s worth benchmarking the model and harness together. The weights alone don’t tell you what the system can do.
Read more: fastino.ai/blog/agent-harnes…
Gizmo Wizard retweeted
DeepSeek V4.1-Flash-1017 can now run locally! 🐳
We made 1-bit retain ~88.2% accuracy while shrinking size from 508GB to 250GB (−52%).
Run on a 128GB+ RAM device.
V4.1 Flash 1017 is the strongest open model to date
Guide: unsloth.ai/docs/models/deeps…
GGUF: huggingface.co/unsloth/DeepS…
🚀 DeepSeek-V4.1-Flash-1017 is here!
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.
🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
1/4
Gizmo Wizard retweeted
This was such a GOATED lecture.
If there is a single thing you should watch about speculative decoding, this is the one!!
Here is what you will learn:
- MTP, DFlash, DSpark, DFlash2...
- How a drafter is trained
- How a drafter is actually coupled with its target model, how it reuses its hidden state and so on
- How training a drafter on a bigger model, different from the target model, is actually leading to a higher mean acceptance length
-How you can use efficient embedding in your lm_head by clustering semantically similar tokens to get extra speed
- How many tokens you can expect the drafter to generate and how to choose that
- How to actually use a drafter via llama.cpp
- And so much more!!!
The speakers are doing such a great job at keeping things simple and educational for everyone, love it!!
Please do yourself a favor and watch the replay if u missed it!
Learn about how to get multipliers in local LLM speed-ups using speculative decoding in this livestream with @mervenoyann, @MaartenGr and @_EldarKurtic
youtube.com/watch?v=H4azmWcq…
Gizmo Wizard retweeted
Great work from @XunzhuoLiu and the vLLM Semantic Router team on D3. All six models, from 0.6B to 27B, now have merged vLLM recipes!
D3 ranked #1 on both Jev Decision Index 0.3.2 boards: 66.3 text and 71.3 vision (Oct. 12 snapshot).
Run multimodal decisions on your own infrastructure through /v1/systemone—with control over data residency, latency budgets and governance.
vLLM serves decision models too: the full D3 family, Clef-Flash, Clef-Omni and Liquid d1, through the System One API.
D3 returns choices, yes/no answers and scores with probabilities over text, JSON, images and videos. Useful for routing, classification and scoring without generating a long response.
In the recipe benchmark, D3 27B delivers 1.8× the requests/sec of its checkpoint server on 1× MI325X at 64 concurrent clients, using the same mixed workload.
Start with a current vLLM nightly:
vllm serve vllm-sr/d3
Recipes for all six sizes:
github.com/vllm-project/reci…
API and supported models:
docs.vllm.ai/en/latest/model…
Introducing Decision 3.0: towards open multimodal foundation decision models ⚡
Open and state-of-the-art from 0.6B to 27B: #1 on both Jev Decision Index boards in our evaluation, on a new Pareto frontier.
HF collection 🤗: huggingface.co/collections/v…
#OpenSource #Jev #LLM
Gizmo Wizard retweeted
You want to teach a robot to put a toy in a drawer. You demonstrate it smoothly. The robot tries — and grabs slightly off, pushing the toy aside. The position has shifted, but the robot keeps going as if nothing changed. More data might help, but collecting it on a real robot is slow and expensive. What if you could give the robot extra training without ever running it on hardware?
Dong Hao's team at Qiyuan Robotics presents HIL-UMI (ArXiv 2609.20659).
HIL-UMI collects experience on a handheld device, then systematically compares what the human does versus what the current model would do — flagging exactly where more demonstrations are needed. When the difference exceeds a threshold, the system prompts the operator to keep recording from that state until the step is complete. No robot execution is required during data collection. Tested on π0.5 with a Franka arm across four tasks, both SFT and HIL-UMI start from the same average score of 44.3. After three rounds, SFT reaches 58.0; HIL-UMI without the advantage mechanism reaches 88.0; and the full method climbs from 62.3 to 94.8. Paper and project page are open.
Result: HIL-UMI trains robots with handheld-only data — no hardware needed during collection. Three rounds lift average score from 44.3 to 94.8 (full method), vs 58.0 for standard SFT, on π0.5 with Franka across four tasks. Open paper and project page.
HIL-UMI: Human-in-the-Loop Data Collection for Robot Learning Without Hardware
Paper: arxiv.org/pdf/2609.20659
Project: hil-umi.github.io/
Our report: jiqizhixin.com/articles/2026…
📬 #PapersAccepted by Jiqizhixin
Gizmo Wizard retweeted
FIRM: Fisher-Informed Retention Merging for Federated LLM Unlearning
Reeyansh Shah, Maram Hasan, Subhasis Chaudhuri, Biplab Banerjee
arxiv.org/abs/2610.21720 [𝚌𝚜.𝙻𝙶 𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙲𝙻]
ALT Federated LLM unlearning addresses targeted knowledge removal when forget and retain data remain distributed across private clients and are inaccessible to the server. The resulting locally trained adapters encode conflicting objectives, making their aggregation fundamentally different from conventional federated optimization. A key failure mode is magnitude explosion, whereby unnormalized cross-client aggregation increases the merged update magnitude and degrades retained utility under stronger unlearning. We propose FIRM (Fisher-Informed Retention Merging), a single-round framework for stable and retention-aware server-side aggregation. FIRM uses client-local Fisher statistics as a unified sensitivity signal to construct a Fisher-normalized base merge and impose explicit control over retain-sensitive parameters. Its second-order retention projection admits a unique minimum-distance solution, while the corresponding dual scales with the number of retain clients rather than the model d
Gizmo Wizard retweeted
Architecting for Hardware Diversity: Meta's End-to-End Co-Design for High-Performance AI Training on AMD GPUs
Meta shows co-designing models and software for AMD GPUs yields up to 1.5x higher training throughput than direct ports from NVIDIA hardware.
arxivb.org/abs/2610.22230
Gizmo Wizard retweeted
$ASTS 🛰️ Europe’s mobile operators want a stronger say in the future of satellite connectivity, with Deutsche Telekom, Orange, Telefónica and Vodafone signing a Letter of Intent to explore a joint bid for Mobile Satellite Services spectrum.
The ambition is to accelerate direct-to-device services, extend coverage into underserved areas and strengthen network resilience, while keeping key infrastructure and governance in Europe. Several strategic points stand out:
🤝 Collaboration could give them greater bargaining power and a clearer route to scale. Combining their customer reach and network expertise should help integrate satellite connectivity into existing mobile services, while reducing fragmentation across European markets.
📱 The proposed model would give all operators open, non-discriminatory access. That commitment will be crucial in determining whether this becomes a genuinely broad platform for competition and consumer choice.
🇪🇺 European sovereignty is central to the announcement, with plans to establish ground infrastructure, coordination and monitoring functions in Europe. The choice of satellite partner, which remains open, will be critical to delivering those ambitions.
💰 The four operators do not plan to invest directly in a satellite constellation, leaving that role to a partner. This pragmatic approach extends their capabilities while managing capital exposure, though the commercial terms will determine how much value they capture.
Existing satellite partnerships remain in place, and this is an exploratory agreement rather than a confirmed spectrum bid or service launch.
The opportunity is to make satellite coverage a seamless extension of mobile connectivity. Success will depend on execution, device compatibility, and pricing that gives customers a compelling reason to pay, along with reassurance that they can stay connected in more places.
Gizmo Wizard retweeted
Great to see more efforts to track auto-research papers!
Curious to keep up to date with auto-research papers?
Just added a category for it on Papers with Code!
Find it here: paperswithcode.co/tasks/auto…
Gizmo Wizard retweeted
When over Poland, a British Airways plane fails to respond to ATC for more than 15 minutes, Poland scrambles fighter jets.
And even if the pilot eventually responds, the Polish Air Force pilot will still come to visually check the cockpit.😬
.
.
Gizmo Wizard retweeted
turns out LLMs are capable of innate embodiment???
arxiv.org/abs/2606.19535 They backdoored an LLM so that it writes insecure code when runing on specific GPUs!
Gizmo Wizard retweeted
GroundAnything
Why locate objects one token at a time? This 4B diffusion model generates grounding predictions in parallel, delivering a 4.51× speedup while setting a new SOTA across 30 benchmarks.
Gizmo Wizard retweeted
The key idea is to train with both an RL loss and an action-prediction loss: RL optimizes for long-term outcomes, while action prediction provides dense supervision at every step.
A simple but very promising direction for improving credit assignment and sample efficiency.
🤔️How to build sample-efficient RL algorithms for long-horizon language agents?
🚀Our answer: Just ONE additional loss on top of GRPO can achieve up to 2× sample efficiency across 5 long-horizon agent benchmarks!
By adding a multi-step action forecasting loss, we encourage the agent to learn action chunks from successful rollouts. Surprisingly, this simple idea works really well!
GitHub: github.com/ulab-uiuc/CoPE
Project page: ulab-uiuc.github.io/CoPE/
Paper: arxiv.org/abs/2610.20597
Here's the thread 🧵👇
Gizmo Wizard retweeted
Confirmed 24K context on iPhone with both K/V on int8. Delta net stays FP16. Decoding also peaks at 70 tok/s with easy to predict context.
Apple Neural Engine on A20 is fast 🚀
Qwen 3.8- 27B on iPhone 18 Apple Neural Engine.
Added draft model , dflash-2 at ~ 2.8 bpw ( mixed 2-4 vector quantization with small drop in acceptance rate)
Base is @PrismML Bonsai-2, 1.585 bpw requantized to 1.5 bpw ( VQ 6 bits per 4 weights)
Decoding speed is 15-60 tps. Fixed context size 16K.
We might push it to 24K.
6GB to 7GB iOS App ceiling is the main limiting factor.
I like this as PoC .
Any suggestions for actual iPhone App on ANE?
Gizmo Wizard retweeted
A coding agent can hand repository search to a small local model.
ColGrep-Agent searches code and returns relevant files and line ranges to your main coding agent.
In LightOn’s 52-question test, delegation cut API tokens by 34%. Average time rose from 13.0 to 21.6 seconds. Accuracy moved from 83.3% to 80.8%, so token savings alone don't settle the trade-off.
The code and weights are Apache 2.0, with local CPU/GPU support. Your main agent's API charges still apply, and selected code may still reach its cloud provider.
A practical experiment for teams: measure search cost, waiting time and missed context on your own repository.
huggingface.co/blog/lightona…
Gizmo Wizard retweeted
That’s the CVT smoke from the LEAP engine, which typically occurs when the engine is operated in hot conditions.
The Center Vent Tube (CVT) functions as a venting mechanism within the air exhaust system and is designed to ventilate the oil. Oil accumulation and coking inside the CVT is a normal design-related phenomenon that can potentially lead to smoke. Such smoke is usually observed after engine shutdown, when the lack of ventilation causes higher temperatures in the CVT during the soak-back period.
In exceptional cases, this can even produce a small flame. It remains completely safe, however, as the flame is contained and the materials in this area are specifically selected to resist fire. The flight crew always has the option to treat it as a tailpipe fire and take remedial action if desired. A fire or flame in the CVT is not considered a safety issue and is expected to self-extinguish without any intervention.
Gizmo Wizard retweeted
Inspired by Nick Cammarata’s interpretability essay “Neural Networks Are Already Clean”, I turned a couple of its ideas into speedups for pretraining without backpropagation. Still in the works. 2.3x speedup, cool xitter.cf/industriaalist/status/…
Replacing backprop with zeroth-order optimization is inevitable.
Introducing Noise: a speedrun for pretraining without backprop.
Last week we presented Dust, the first method to approach backprop with SGD for pretraining transformers. We think another 1,000x improvement on top of it is possible, and it will require *co-designing* credit assignment, optimizers, architectures, and eventually, hardware.
We built Noise to iterate on this faster.
Gizmo Wizard retweeted
Another style of pagoda garden. Built with plan mode on Qwen3.8-27B EXL3 6.0, all running locally on my DGX Spark.
TandemLLM has a built-in browser that previews your app in a sandbox container right on the Spark.
Start the chat from your phone, close the browser, come back later and pick up where you left off, just like Claude Code 🔥
Video below is straight from TandemLLM 👇
