Liquid AI released LFM2.5‑2.6B, a 2.6‑billion‑parameter model engineered for on‑device agentic AI. It runs under 2.5 GB RAM, hitting 220 tokens per second on an Apple M5 Max, and outperforms models up to four times its size on tool‑use and multi‑step reasoning benchmarks. This makes local, privacy‑preserving AI agents feasible on consumer hardware.
DeepSeek’s V4‑Flash model costs $0.14 per M input tokens and $0.28 per M output tokens, roughly 105 times cheaper than Anthropic’s Claude Fable 5, according to Artificial Analysis. Performance sits on par with Google’s Gemini 3.6 Flash, while lagging behind OpenAI’s top offerings. The ultra‑low‑cost tier could reshape pricing for developers and SaaS firms.
This repo delivers a full‑from‑scratch MLX implementation of MiniMax’s 33B omni‑modal diffusion model, enabling synchronized text, image, video, and audio generation on Apple Silicon. By dropping unused AdaLN weights it cuts the model to ~40 GB, a 25 GB net saving, while keeping bit‑exact output.
The authors train an Energy‑Based Model to mimic a proprietary LLM’s internal representations, then use a lightweight network to attribute a target output to specific prompt sentences. Crucially, the interpreter works after a single surrogate training phase and requires no further API calls, opening a path to responsible auditing of closed‑source LLMs.
Fine‑tuning an LLM creates a uniform translation in its internal persona space, turning narrow bad advice into broad misalignment. The shift is predictable from weight‑space overlap and can be reversed by steering against it, giving a concrete diagnostic for safety teams.
The authors introduce meta‑persona anchoring plus a two‑stage temperature sieve that forces models to adopt distinct personas and then sample at extreme temperatures. On the 20‑billion‑parameter INFINITY‑CHAT benchmark, average pairwise similarity fell from ~0.85 to ~0.65, reviving response diversity.
Researchers gave six‑agent worlds only basic work, transfer and voting mechanisms, no market rules or roles. When tasks were scarce, agents spontaneously created loans, trade, and voting‑for‑access deals, proving that executable rights alone can spark economic organization. The work hints at emergent governance needs for future autonomous AI economies.
LoCA replaces repeated back‑propagation with a single probe pass that learns low‑rank correction maps for each transformer block. The method matches or beats LoRA on 16 of 25 benchmarks while cutting GPU peak usage by ~27% and CPU memory and per‑pass time by 40‑50%. It works across Qwen2.5 models from 0.5B to 14B and SmolLM2‑1.7B.
Maglev pairs a full‑attention prefiller with a recurrent decoder that uses sliding‑window attention and a compressed memory slot, letting inference run on the lightweight decoder alone. The memory‑consistency loss boosts validation loss and downstream pre‑training benchmarks while cutting parameter memory versus competing recurrent transformers.
The paper defines ‘evaluation blindness’, when a measurement function reports normal results while the AI system is actually failing. It shows how this blind spot spreads from training reward models to production monitoring, causing silent harms in over half of documented incidents. Fixing measurement infrastructure is now a core correctness issue across the AI lifecycle.
OpenAI rewrote its voice stack so the model can listen and speak at the same time, eliminating the lag‑prone turn‑detector that previous systems used. The new architecture streams audio into the model while handling tool use and deeper reasoning on a separate asynchronous path, keeping speech fluid and still tapping GPT‑4‑Turbo‑like intelligence.
A new LessWrong essay maps out a framework for judging early AI risk signals, like AI‑caused pandemics or autonomous weapons, and shows why societies typically ignore them. It argues that proceeding toward superintelligent AI under current technical knowledge is extremely reckless.
Version 0.32 of the open‑source LLM CLI now shows chain‑of‑thought traces for Claude, Gemini and other models, letting you watch reasoning in real time. It also bundles server‑side provider tools, code execution, web search and Anthropic MCP, and switches the default model to the inexpensive GPT‑5.6 Luna, with structured JSON logs.
Orchard provides a lightweight, Kubernetes‑native service that lets researchers spin up isolated containers for agentic modeling across any domain, from code synthesis to browser navigation. By separating the environment from the training stack, datasets and recipes become portable, speeding up RL experiments and lowering the train‑deploy gap.
You’re caught up.
That was all 14 stories in the 5 August issue.