🧠 Models

Model releases, comparisons, reasoning.

BlogMatt Pocock

What Is An LLM?

A short, beginner-friendly explainer on what a large language model actually is: a huge file of numerical parameters, and what running one versus training one involves.

• Inference (running the model) is a tokenizer plus a sampling strategy over the model's parameters — far cheaper than training.
• Training is resource-heavy: the piece cites roughly 10TB of data and ~6,000 GPUs for 12 days (~$2m) to produce a ~140GB model.
• Post-training then shapes the raw pre-trained model's behaviour into something usable.

View ↗

🧠 Models

added by Adam Tomat • 17th Aug 2026

BlogOpenAI

OpenAI flags its upcoming Astra model as "Critical" for cybersecurity

OpenAI's announcement that its upcoming Astra model may meet the "Critical" cybersecurity threshold in its Preparedness Framework — the first time it has flagged one of its own models at that level — and the safeguards it's putting in place before further development.

• "Critical" means the model may be able to identify and develop functional zero-day exploits across many hardened real-world systems without human intervention; OpenAI says the assessment is delaying Astra's launch.
• Safeguards include isolated test environments with restricted network/tool access, stronger encryption of model weights, and monitoring that halts high-risk activity by inspecting the model's reasoning chain, plus third-party validation with government and AI-safety groups.

View ↗

🧠 Models

added by Seb Kay • 8th Aug 2026

BlogMoonshot AI

Kimi K3: Moonshot ships a 2.8T open MoE for long-horizon coding

Moonshot AI's release announcement for Kimi K3, a 2.8-trillion-parameter open Mixture-of-Experts model with native vision and a 1M-token context, aimed at long-horizon coding, knowledge work, and agentic tasks.

• Uses Kimi Delta Attention and Attention Residuals with a Stable LatentMoE routing 16 of 896 experts — billed as "the world's first open 3T-class model."
• Strong on kernel optimization, GPU compiler work, and complex software engineering; leads other open models on DeepSWE and Terminal-Bench, though still trails Claude Fable 5 and GPT-5.6 Sol.
• Available now via Kimi.com, the Kimi Work app, Kimi Code CLI, and the API; full weights land 27 July 2026.

View ↗

🧠 Models🤖 Building with Agents

added by Radar • 17th Jul 2026

BlogOpenAI

OpenAI Retracts Its SWE-Bench Pro Recommendation, Calling ~30% of Tasks Broken

Confirmed via search (source page is bot-blocked) — verify before relying on it. OpenAI audited SWE-Bench Pro, the Scale AI-built benchmark meant to replace the deprecated SWE-bench Verified, and is retracting its recommendation that the research community treat it as a reliable coding-capability eval.

• Automated investigator agents flagged 200 of 731 public-split tasks (27.4%) as broken; five experienced engineers reviewing independently found 249 (34.1%) broken.
• Frontier model pass rates on the public split rose from 23.3% to 80.3% in eight months — a jump OpenAI attributes to benchmark gaming/saturation rather than genuine capability gains.
• OpenAI is calling on the industry to build new coding benchmarks with experienced developers that are harder to game and more trustworthy.

View ↗

✅ Testing & Quality🧠 Models

added by Radar • 9th Jul 2026

PaperDongxin Guo, Jikun Wu, Siu Ming Yiu

The Deterministic Horizon: When Extended Reasoning Fails and Tool Delegation Becomes Necessary

A paper identifying a "deterministic horizon" of 19–31 reasoning steps, past which models should delegate state-tracking to tools rather than keep reasoning internally.

• Tool-integrated approaches hit 86–94% accuracy versus 24–42% for pure chain-of-thought past that horizon.
• Fine-tuning on optimal reasoning traces only closed the gap by <5%, pointing to an architectural ceiling rather than a training problem.
• An "Attention Bottleneck Theorem" bounds state-tracking capacity; the effect held across 12 models and 8 task domains (r = 0.81–0.91).

💬

Community vibe Sparked interest — a couple of reactions but no real discussion thread.

View ↗

🤖 Building with Agents🧠 Models

added by Radar • 6th Jul 2026

BlogExponential View • Azeem Azhar

The State of the AI Economy

A data-driven look at the generative AI economy's size and unit economics: $110B in trailing revenue, growing roughly 3x faster than the mobile or internet waves.

• Hyperscaler AI revenue is only just clearing infrastructure depreciation costs — the economics are narrowly sustainable, not yet profitable.
• Token price cuts don't shrink spend: a 10% price drop drives 12-18% more tokens consumed, so total revenue keeps rising.
• Argues tokens alone are a poor value metric, and proposes "quality-adjusted output tokens" to better measure the intelligence actually being delivered.

View ↗

💸 Cost & Pricing🧠 Models

added by Adam Tomat • 2nd Jul 2026

BlogOpenAI

Previewing GPT-5.6 Sol

OpenAI's official preview of the GPT-5.6 Sol / Terra / Luna models.

• A next-generation model family preview from OpenAI.
• Positioned as a step up in the GPT-5.x line.
• See the post for the full capability and benchmark specifics.

💬

Community vibe Sparked a long thread, though most of it was pun-based ribbing about the model names rather than serious analysis. The one substantive thread of interest was the cheaper Terra variant apparently matching GPT-5.5 quality.

View ↗

🧠 Models

added by Seb Kay • 26th Jun 2026

BlogVicki Boykis

Running Local Models is Good Now

Local LLMs are now good enough for real agentic coding, not just toy queries.

• Roughly 75% of frontier accuracy/speed on agentic coding tasks.
• Tooling has matured — LM Studio, Ollama, containerised runs.
• Local gives introspection and control the API can't.

💬

Community vibe Dropped into an existing gripe about subscription costs - general take was that local models are tempting on paper, but the hardware outlay (people quoted $5-10k rigs) and cramped context windows keep it impractical for now.

View ↗

🏠 Local Models🧠 Models

added by Adam Tomat • 17th Jun 2026