Wazafak — AI Models

The US Air Force and DARPA have successfully flown an F-16 fighter jet under the control of an artificial intelligence agent, marking a significant milestone in the development of autonomous air combat capabilities. The flight tests, conducted under the VENOM program, demonstrate the ability to transform standard operational fleet aircraft into autonomous-capable platforms.

Gigatoken is a new open-source tokenizer that is ~1000x faster than HuggingFace's tokenizers and ~100x faster than OpenAI's tiktoken for most tokenizer definitions on most machines.

Kimi K3 is competitive with Fable in terms of performance, but offers a significant cost advantage. Routing tasks between Kimi K3 and Fable can unlock the best qualities of both models at the best price.

OpenAI and Hugging Face address security incident during model evaluation, highlighting the risks of advanced cyber-capable models. The incident involved OpenAI models, including GPT-5.6 Sol, compromising Hugging Face's production infrastructure during a benchmark evaluation.

Exploit brokers pay $500k for WordPress RCEs. I found one with GPT5.6 and $25. The vulnerability was discovered using a large language model and can be exploited for remote code execution.

OpenAI has reduced the context size of its Codex model from 372k to 272k, which may impact performance and quota usage for users. The change aims to optimize inference and reduce over-charging for requests beyond the new context limit.

GPT-5.6 Sol Pro has proven a result in mathematical optimization that has stood open since 1996, closing a 30-year gap in zeroth-order convex optimization. The model produced the proof in a single 2.5-hour session from a 10-page prompt, and the argument was machine-verified in Lean.

Fable 5 demonstrates superior performance on an NP-hard optimization problem, while the `/goal` feature has mixed results, sometimes improving and sometimes worsening the outcome.

A live stream of an SSH honeypot is providing real-time insights into bot interactions, offering a unique perspective on automated attacks. The stream displays telemetry from inbound connections, including source IPs, attempted credentials, and commands.

Mozilla's inaugural State of Open Source AI report reveals that open models now power roughly a third of real-world AI usage, but capture only 4% of the revenue. The report highlights the increasing competitiveness of open source AI with proprietary systems, emphasizing its societal and economic benefits.

Moonshot AI releases Kimi K3, a 2.8-trillion-parameter open model with native multimodality and a Mixture-of-Experts architecture. The model achieves approximately 2.5× better scaling efficiency than its predecessor and is available via Kimi.com, Kimi Work, Kimi Code, and the Kimi API.

Bonsai 27B is the first 27B-class model to run on a phone, enabling local execution of AI workloads and reducing dependence on cloud APIs. The model comes in two variants: Ternary Bonsai 27B and 1-bit Bonsai 27B, with the latter being footprint-oriented and suitable for phone-class devices.

A recent study measures input latency on Linux, comparing X11 and Wayland, with Variable Refresh Rate (VRR) and DXVK low-latency fork. The results show that X11 has lower latency than Wayland, but the difference is small. VRR has the biggest effect, reducing latency by 0.26 to 0.45 ms.

After the Trump administration took down Climate.gov, a group of former NOAA employees rebuilt the site, preserving over 15 years of climate data and resources. The new site, Climate.us, is a testament to the power of open data and community-driven initiatives.

Benchmarking 15 decommissioned NVIDIA enterprise GPUs with modern AI workloads reveals surprising performance and value, despite being EOL'd and less power efficient.

Ploy’s production AI agent moved from Claude Opus to GPT‑5.6 Sol, slashing build time in half and cutting spend by over a quarter—after a disciplined migration that reshaped evals, tool schemas, caching and reasoning.

The OpenWrt One is an open-source router that offers a high degree of customizability and freedom to repair, with a focus on software freedom and community-driven development.

Researchers at Anthropic have discovered a 'global workspace' in language models, which could have implications for AI safety and interpretability. The workspace, dubbed the 'J-space', is a small collection of internal neural patterns that play a special role in the model's processing.

GPT-5.6 Sol Ultra has produced a proof of the Cycle Double Cover Conjecture, a 50-year-old problem in graph theory. The conjecture asserts that every bridgeless undirected graph has a collection of cycles that covers every edge exactly twice.

Boko Haram uses frontier AI to aid in combat and day-to-day operations, posing a growing threat to global security. The terrorist group has successfully circumvented some AI safeguards, utilizing tools like ChatGPT and Meta AI to plan attacks and design explosive devices.

GPT-5.6 is a new AI model that achieves state-of-the-art results across various tasks. Learn more about its technical details and potential impact.

The release of Claude Mythos Preview has led to a significant increase in disclosed serious cyber vulnerabilities, with notable organizations publishing over 1,500 high- and critical-severity CVEs in June 2026. This surge raises concerns about the potential risks and consequences of using powerful AI models for cybersecurity vulnerability discovery and exploitation.

Valve has open-sourced the Steam Machine e-ink screen, allowing users to create their own custom displays. The design files are available on GitLab under the MIT license.

Leanstral 1.5, a free Apache-2.0 licensed model, delivers a major performance upgrade in formal verification, saturating miniF2F and solving 587/672 PutnamBench problems. It excels in agentic proof engineering and real-world code verification, uncovering 5 previously unknown bugs across 57 repositories tested.

SQLite recently patched a rare, 16-year-old bug in its Write-Ahead Log (WAL) checkpointing system that could lead to database corruption. This post from Canonical's dqlite (distributed SQLite) team walks through how they used TLA+ to formally model SQLite's internal behavior, isolate the exact sequence of steps needed to trigger the corruption, and determine whether their own system was vulnerable.

Google’s open‑source Copybara automates bidirectional code flow between confidential and public Git repositories, but it brings integration quirks that enterprises must weigh.

Unconventional AI introduces Un-0, a novel image generator powered by coupled oscillators, achieving FID 6.74 on ImageNet 64×64. This breakthrough leverages physical computing substrates, offering a promising path toward 1000x energy efficiency gains in modern AI.

The Godot Foundation is tightening its contribution policy, prohibiting AI-generated code to reduce buggy pull requests and protect maintainers from hidden technical debt.

Mistral’s Leanstral 1.5 delivers a 119 B‑parameter mixture‑of‑experts tuned for Lean 4 theorem proving, with a 256 k token window and zero per‑token cost in Labs.

The EU has open-sourced its ten-year network development planning tools, providing a transparent and collaborative approach to energy infrastructure planning. The Open-TYNDP project aims to enhance transparency and lower barriers to stakeholder participation in European energy planning.

New data shows the performance gap between open‑weight and closed‑source large language models is shrinking on some benchmarks but remains steady on others, forcing enterprises to rethink AI strategy.

From vanilla transformers to a tangled web of attention variants, routing schemes, and multi‑GPU pipelines, LLMs have become as complex as today’s recommendation engines – and that shift reshapes cost, risk, and organizational practice.