AI and infrastructure news
Filter by type: Post307
- 1 day ago x.com
Dessa: full-duplex speech-to-speech model announced for consumer GPUs
A post announces Dessa, a full-duplex speech-to-speech model that, according to its author, was trained only on speech and runs on consumer GPUs. Relevant to locally hosted conversational voice.
- 1 day ago x.com
Open TTS Leaderboard adds 31M and 35B models and the UTMOS metric
The Open TTS Leaderboard update compares a 31M-parameter open model with a 35B one on English WER and adds the UTMOS metric for speech naturalness.
Author argues short-lived bearer tokens need proof of possession
The post argues that OIDC bearer tokens with 15-minute expiry do not prevent intrusions, and that proof of possession, such as TLS binding, is the only real solution. The author also predicts OIDC will end soon.
- 2 days ago x.com
Meta paper proposes byte-level models without a tokenizer
Meta's paper 'Breaking the Token Ceiling' describes models trained on raw bytes with distillation. A post credits 1B byte models with performance equal to or above Llama 3.2-1B and Gemma-3-1B, claims still to be confirmed.
Private CA roots can issue certificates without transparency log entries
The post claims that TLS inspection roots installed on corporate laptops can issue certificates for any domain without appearing in certificate transparency logs, a trust gap that calls for local monitoring.
- 4 days ago x.com
Ant Ling releases five Ming-3.0 speech models with open weights
Ant Ling announced the open-source release of five Ming-3.0 models covering speech synthesis variants, audio understanding, dialogue and automatic speech recognition for dialects.
- 4 days ago x.com
Streaming ASR: decoding speech without seeing future audio
The Tongyi Speech AI post explains streaming speech recognition, where words must be committed while the speaker is still talking. Latency and lookahead shape the model and decoder design for live voice applications.
- 4 days ago x.com
Kyutai releases smaller Pocket TTS versions with up to 1.9M parameters
Kyutai Labs has announced smaller Pocket TTS versions at 18.2M, 6.9M and 1.9M parameters, and claims the 1.9M model runs 103x faster than real time on a single CPU core, with voice cloning.
- 4 days ago x.com
Open source TTS models smaller than Kokoro, with voice cloning
The post announces open source text-to-speech models about half the size of Kokoro (82M parameters), with claimed gains in naturalness and accuracy and support for voice cloning. They could lower the cost of running speech synthesis locally.
- 4 days ago x.com
Phonon: 40M-parameter multilingual TTS model for locally run voice
According to the author's announcement, Phonon is a text-to-speech model of about 40M parameters, with multiple languages, a voice library, and local use through the speech-to-speech package.
- 4 days ago x.com
Phonon: open 31M-parameter TTS running in the browser via WebGPU
Post reports a port of Phonon, an open 31M-parameter text-to-speech model from GradiumAI, running fully in the browser via WebGPU in half precision, with the author citing about 47x real time on an M4 Pro and 7x on a Phone 13 Pro.
- 5 days ago x.com
Diffusion text models generate the full output at once: what changes for inference
Post describes diffusion text models that produce the whole response simultaneously, instead of token by token like autoregressive LLMs. According to the author, generation is fast, with a latency and cost profile distinct from standard inference.
- 5 days ago x.com
Intel GPU D270 (Crescent Island): inference accelerator with high-capacity LPDDR5X memory
A post describes Intel's D270 GPU (Crescent Island), a PCIe 5.0 card for inference with 32 Xe cores, XMX matrix engines and about 160 GB of LPDDR5X, expandable to 480 GB.
- 6 days ago x.com
AuroraASR-1M: offline speech recognition on CPU with a 1 MB model
The post describes a 949K-parameter Conformer-CTC model for English, 1.06 MB in INT8, Apache 2.0 licensed, running at about 436x real time on one CPU thread with an 11.43% WER on test-clean.
Local LLM benchmarks on a Ryzen AI Max+ PRO 495 and the decision to return the machine
Throughput analysis of four large open-weight models run with llama.cpp/Vulkan on a 192 GB unified-memory APU, reporting 18 to 58 tps and a return made for cost reasons.
- 6 days ago x.com
Pruning 6 layers of Qwen3-8B-Base with LaCo cuts the model by 16.7%
The post reports removing up to 10 layers from Qwen3-8B-Base using LaCo. With 6 layers removed, the model is 16.7% smaller and, according to the author, keeps 78-94% of its reasoning scores, with a perplexity ratio of 1.68x.
- 15 days ago x.com
Artificial Analysis: HeyGen Voice tops controlled voice TTS ranking
Artificial Analysis ranked HeyGen Voice first in the Controlled Voice TTS Arena, with an Elo of 1,201 across 8 cloned voices in US and UK English, ahead of competing commercial voice models.
- 15 days ago x.com
Social media post on RoboJEPA: predictable scaling in robotic world models
A social media post describes RoboJEPA, an 8B-parameter JEPA world model trained on 15,022 hours of robot video, and scaling-law fits that anticipate the results of the 4B and 8B models.
- 16 days ago x.com
Voice clones of non-native English speakers rated more human: what it means for companies
A NeurIPS 2026 paper cloned the voices of 86 non-native English speakers using three voice-synthesis systems. In 4,000 human ratings, the clones were judged warmer, more authoritative and more human, which raises trust risks for voice channels.
DeLM: decentralized code agents with shared context, without a central coordinator
Stanford's article presents DeLM, which replaces a central coordinator with shared context and a task queue. According to the author, it is up to 2.49× faster and more accurate than the Claude Code and Codex baselines.
- 16 days ago x.com
Samsung opens LittleBit, a method that compresses 13B LLMs to under 1 GB
According to the post, Samsung's LittleBit uses latent factorization and sub-1-bit levels to shrink the weights of 13B LLMs to under 1 GB. Worth evaluating for local deployment.
- 17 days ago x.com
Quantizing 10M EmbeddingGemma vectors to 1 bit with MRL cuts RAM from 30GB to 0.4GB
The author's post reports compressing 10 million EmbeddingGemma embeddings to 1 bit, with MRL reduction to 256 dimensions, dropping RAM from about 30GB to 0.4GB at a stated quality loss of about 5%.
- 17 days ago x.com
Single-voice 9MB TTS model distilled from Kokoro-82M for local voice output
A post describes a single-voice, single-language text-to-speech model of about 9MB of weights, distilled from Kokoro-82M and aimed at local voice output on resource-constrained devices.
Prompt injection vectors: how to test and harden AI systems
The post gathers prompt injection vectors and strategies with examples, useful for assessing and hardening the security of applications built on language models.
Auditable self-improving agents
An article published on April 19, 2026 proposes a protocol for agents to suggest, evaluate, and implement improvements with auditable traceability and rollback.
Code agents: evaluating safety and performance
Microsoft reports a 14-billion-parameter coding agent scoring 45.3% on SWE-Bench Verified, and a training method that accounts for tests unintentionally broken while adding features.
Recursive agents exchange latent states to collaborate
A May 1, 2026 publication describes a paper on agents refining shared latent states and reports more precise collaboration, faster execution, and lower token use in experiments.
- 17 days ago x.com
Fine-tuning a 14B code model with SFT and DPO
A January 20, 2026 article describes tuning a 14B model on code conversations and reports figures, costs, and difficulties.
- 17 days ago x.com
Architecture also shapes continual learning
An article published on January 17, 2026 reports that wider networks and less global average pooling may improve retention, while ResNets and WideResNets learn new tasks quickly but forget more.
- 17 days ago x.com
Private evaluation data to reduce ASR test leakage
On May 6, 2026, the Open ASR Leaderboard added private evaluation data from Appen and DataoceanAI to reduce test-set contamination and overfitting.
- 17 days ago x.com
Benchmark reports 2.89× faster TTFT at 64K context
Published on May 9, 2026, the benchmark reports that Luce PFlash achieved 2.89× faster time to first token than Ollama with a 64K context.
- 17 days ago x.com
Benchmark tests retrieval in legal RAG systems
A paper published on March 7, 2026 presents a criminal-law retrieval benchmark with 4,876 passages and 100 complex questions written by legal experts.
- 17 days ago x.com
BM25: lexical search without embeddings
A post published on February 9, 2026 explains how BM25 scores documents using term rarity, diminishing returns for repetitions, and document-length normalization.
Bootloaders: assessing risks and defenses
An article published on June 5, 2026, covers attack surfaces, vulnerability detection, and defenses in firmware, operating systems, and monolithic bootloaders.
- 17 days ago x.com
CALM proposes predicting vector blocks, not tokens
A text published on January 24, 2026 describes CALM models that predict blocks of meaning, claiming 4× more information per step and 44% less training compute.
- 17 days ago x.com
CLI estimates VRAM before loading models
The do-i-have-the-vram tool estimates the GPU memory needed to load a model without loading it, helping assess whether it fits in available VRAM.
- 17 days ago x.com
How attention and MLPs store knowledge
Published on April 22, 2026, the paper examines trade-offs in knowledge storage by attention matrices and MLPs, identifying limits linked to finite samples and compute, as well as noise in long contexts.
- 17 days ago x.com
How to evaluate voice models for products and operations
On November 6, 2025, Artificial Analysis reported that Inworld TTS 1 Max led Speech Arena, a blind comparison of text-to-speech models based on human evaluations.
- 17 days ago x.com
How to compare MTP inference on older GPUs
A May 16, 2026 publication reports llama.cpp throughput benchmarks with MTP for Qwen 3.6 models on three GTX 1080 Ti GPUs and describes a setup that can inform internal testing.
Building a debugger from scratch: mechanisms and applications
A series of articles published on June 3, 2026, explores building a debugger, from connecting to a process to handling source code and symbols.
Let's assess your project
I reply with scope and a proposal.