Fable 5 Returns, Gets Hacked Cheaper by Nightfall

July 5, 2026

28 topics · 25 sources

AI Models
AI Search Simon Willison AICodeKing

Claude Fable 5 Returns Globally — Restricted, Pricey, and Already Being Hacked Around

Anthropic re-released Claude Fable 5 worldwide after the US government's ban lifted, but with heavy new restrictions: subscription access ends July 7, and usage is capped at 50% of weekly limits, so routine coding and debugging now silently fall back to the dumber Opus 4.8 [1]AI Search. The return played out from three angles today: a news report on the restrictions, a maintainer's real-world bill, and a community tool built to cut that bill down.

Read more

The return, and what got nerfed

Fable 5 launched weeks ago based on the allegedly "too dangerous to release" Mythos model, then was banned by the US government days later. After Anthropic implemented new safety classifiers and coordinated with the government, the ban lifted and Fable 5 came back globally ~14:11 [1]AI Search. The new safety classifier reportedly blocks over 99% of jailbreak attempts but causes more false positives, and the version that returned is "more castrated" than before: coding and debugging tasks fall back to Opus 4.8, undercutting Fable's main agentic-coding use case. Benchmarks show it hallucinates slightly less but performs roughly half as well on debugging/refactoring; Arena leaderboard confidence intervals mostly overlap with rivals, showing no statistically significant edge.

A maintainer's real bill: $149.25 and five release-blocking bugs

Simon Willison used Fable 5 as a coding agent to review and largely write the sqlite-utils 4.0rc2 release ahead of the model's July 7 subscription deprecation [2]Simon Willison. Fable flagged five "release blocker" issues, the worst being a data-loss transaction bug where delete_where() never committed and silently poisoned the connection for subsequent transactions. A cross-model review with GPT-5.5 — now routine practice for Willison — turned up two more Priority-1 bugs: db.query("update...") auto-committing before raising a ValueError, and INSERT...RETURNING only committing after the iterator was fully exhausted, contradicting the docs. The full session spanned 37 prompts across 30 files (+1,321/−190 lines) and cost $149.25 measured via AgentsView — $141.02 for the main session plus a few dollars in specialized review agents. Willison noted he was able to check in on the agent occasionally while attending a 4th of July parade, and that release notes are exactly the "boring, predictable, accurate" writing he's comfortable outsourcing.

The community response: Caveman cuts Fable's chatty output by 65-75%

AICodeKing frames the core complaint plainly: Fable 5 is widely considered the best coding model available, but at $10/M input and $50/M output tokens — literally double Opus 4.8 — and a habit of writing long, chatty explanations, users pay premium prices for filler [3]AICodeKing. Caveman, a ~69,000-star GitHub project, installs as a Claude Code (and Codex/Gemini CLI/Cursor/~30 other agents) skill in one line and forces terse "caveman speak" responses — "New object ref each render. Inline object prop equals new ref equals re-render. Wrap in memo" instead of a full paragraph — cutting 65-75% of output tokens across a 10-prompt benchmark (22-87% range) while preserving technical accuracy, since it only trims narration, not Fable's actual thinking/reasoning tokens. A /caveman compress command also rewrites CLAUDE.md/project notes into dense caveman-speak, cutting ~46% of input tokens on every future session. The host also cites an unnamed March 2026 paper finding that constraining models to brief responses improved accuracy by 26 points on certain benchmarks.

"This new version of Fable 5 is even more castrated." — AI Search
"Caveman no make brain smaller. Caveman make mouth smaller." — AICodeKing
Tools: Claude Fable 5, Mythos, Opus 4.8, Caveman, Claude Code, sqlite-utils, GPT-5.5, AgentsView
AI Models
AI Search

Claude Sonnet 5 Panned as Overpriced and Underpowered

Anthropic's smaller Claude Sonnet 5 landed to sharp criticism: it costs more than double GPT-5.5 while scoring behind it — and behind the older Opus 4.8 — on Artificial Analysis's intelligence index, only barely ahead of open-source GLM 5.2 [1]AI Search.

Read more

On the intelligence-versus-cost chart, Sonnet 5 is simultaneously more expensive and less intelligent than both GPT-5.5 and Opus 4.8, and it's also slower than GLM 5.2 ~18:12. Even at extra-high to max thinking levels it doesn't undercut Opus 4.8 on price. The presenter's conclusion: "there's absolutely no reason to use this over existing models."

"This new Sonnet 5 is incredibly expensive. It costs more than double the price of GPT 5.5 even though it's dumber than GPT." — AI Search
Tools: Claude Sonnet 5, Opus 4.8, GPT-5.5, GLM 5.2, Artificial Analysis
Hot Take
AI News & Strategy Daily | Nate B Jones

Nate B Jones: AI's Real Bottleneck Is Imagination, Not Execution Cost

As AI execution gets cheaper, everyone's output is converging rather than differentiating — Nate Jones argues this isn't a tooling problem, it's proof that value has moved from execution to imagining what to ask AI to do in the first place [4]Nate B Jones.

Read more

The $9 vs. $40 experiment

HashiCorp co-founder Mitchell Hashimoto benchmarked three model tiers on routine coding work: a budget model (GLM 5.2) finished in minutes for under $1, GPT-5.5 cost about $1.50, and a frontier model cost $9 for 40 minutes — yet all three produced equally acceptable output ~02:02. But on a second test — optimizing gnarly systems code that didn't exist on any backlog, a problem Hashimoto invented because he suspected new capability had emerged — the frontier model took 2 hours and cost $40, reaching a level of optimization Hashimoto, an expert in the domain, said he couldn't hit alone.

Execution vs. imagination

Nate's framing: cheap execution is a commodity everyone can drive down, while the multiplier comes from targeted frontier-model use on questions that change what gets built at all. He illustrates with Blackberry vs. Apple — Blackberry executed brilliantly within a category everyone had already imagined, while Apple imagined a different category and won, making its execution "worth a hundred times" Blackberry's ~04:04. A factory-electrification analogy makes the same point about organizations bolting AI onto old structures instead of redesigning around it, and Stripe's reported 50-million-line, single-day migration is reframed as the payoff of years of prior investment in verification infrastructure, not a one-day miracle.

The personal and organizational test

Nate's diagnostic: has your AI task list actually changed in the last 12, 6, or 3 months, or are you just running the same list faster and cheaper ~06:07? He also walks through a real hyper-targeted marketing workflow found on X — combining Google Maps geographic analysis, spatial reasoning to find sun-exposed unshaded porches, and 3D modeling via Blender to generate custom direct-mail offers — as an example of imagination unlocking a workflow no one could have specified in a prompt summary. His closing warning for leaders: you can't hire your way out of an imagination gap with one visionary, since imagination only fires when paired with deep organizational context.

"Whenever execution gets cheap the value doesn't disappear, it just moves." — Nate B Jones
"The ceiling on what AI is worth to you was never the model or the price or the prompt pack... It's the size of your list of things you know how to ask for. It's your imagination." — Nate B Jones
"You can't hire your way out of this with one imaginative person." — Nate B Jones
Tools: Claude Fable 5, GPT-5.5, GLM 5.2, Google Maps, Blender
Hot Take
AI News & Strategy Daily | Nate B Jones

Delegate Whole Jobs, Not Prompts, to Frontier Models

In a companion short, Nate Jones argues that with today's larger models, the right unit of delegation has grown — hand a frontier model something as big as a whole consulting engagement and only review the final output, rather than micromanaging small prompts [5]Nate B Jones.

Read more

He contrasts this with 2023-2024, when "asking big" reliably backfired: models would lose the thread partway through a task, invent sources, and return confidently wrong numbers. The implication is that reliability over long, multi-step tasks has improved enough that the right prompting strategy has shifted from small, verifiable asks toward large-scope delegation.

"Bigger means it can pick up and carry the job and I can trust it to do that and I just have to have a look at it when it's done." — Nate B Jones
Tools: Claude Fable 5
AI Models
AI Search

LongCat 2.0: Meituan Trains a 1.6T Model Without a Single Nvidia GPU

Chinese food-delivery giant Meituan released LongCat 2.0, a 1.6-trillion-parameter open-source MoE model (48B active params, MIT license) trained entirely on non-Nvidia ASIC super-pods — rumored to be Huawei chips — with no training rollbacks or loss spikes [1]AI Search.

Read more

On agentic coding and reasoning benchmarks like Terminal Bench and SWE-bench, LongCat 2.0 beats Google's Gemini 3.1 Pro and edges close to top closed models including GPT-5.5 and Claude Opus ~02:24. It works with harnesses like Claude Code, OpenClaw, and OpenCode, and runs on NPUs as well as GPUs, though the FP8 checkpoint is over 2TB and needs stacked enterprise GPUs to run. The headline significance: training frontier-scale models on non-Nvidia hardware has been largely undocumented until now.

"Their pre-training had no rollbacks or irreversible loss spikes, demonstrating that we have capacity to conduct frontier scale training on alternative hardware platforms." — AI Search
Tools: LongCat 2.0, Meituan, Claude Code, OpenClaw, OpenCode, Gemini 3.1 Pro, GPT-5.5
AI Models
AI Search

Agents A1: a 35B Agentic Model That Beats Trillion-Parameter Rivals

Agents A1 is an open-source 35B-parameter MoE model built specifically for agentic workflows that beats larger models like Kimi K2.6 and the over-trillion-parameter DeepSeek V4 Pro, and even outperforms GPT-5.5 on several science benchmarks [1]AI Search.

Read more

It's especially strong on science benchmarks (Humanity's Last Exam, Frontier Science, SEAL, GAIA) and on instruction following ~08:08. Released with FP8 (37.7GB) and Q4 (21GB) checkpoints that fit on high-end consumer hardware, the presenter suggests it may be the new leading offline-runnable consumer model, beating the previous leader Qwen 35B across the board.

Tools: Agents A1, Kimi K2.6, DeepSeek V4 Pro, GPT-5.5, Qwen 35B
AI Models
Sam Witteveen

MiniCPM-5: Testing Karpathy's "Cognitive Core" Idea at 1B Parameters

Sam Witteveen puts MiniCPM-5 — a 1B dense, Llama-style model from Tsinghua's OpenBMB lab with a 128K context window — through its paces as a real-world test of Andrej Karpathy's "cognitive core" concept: a tiny model that keeps reasoning and tool-use ability while offloading encyclopedic knowledge to lookups [6]Sam Witteveen.

Read more

Specs and training

MiniCPM-5 ships under Apache 2.0 across base, SFT (~400B tokens, half "deep thinking" and half "hybrid"), and a fully RL-trained version with on-policy distillation that boosts math/code/instruction scores while cutting down on overly long responses ~03:02.

Benchmarks: token efficiency and honesty about not knowing

Third-party Artificial Analysis numbers show MiniCPM-5 uses 31x fewer tokens than Qwen3's 0.5B-class reasoning model and 8x fewer than its non-reasoning version for comparable scores ~07:06. On the hallucination-penalizing AA Omniscience benchmark it scores around -1, versus deep-negative scores for peers — evidence it's comparatively good at saying "I don't know" instead of confabulating.

Hands-on: solid at tool use, shaky at instructions and long context

Single and repeated tool calls, multi-step chains, and RAG-style fetch tasks all worked well ~14:14, but long agentic trajectories (~12 sequential tool calls) were hit-or-miss, and GLM-5.2 handled the same test with no trouble by comparison. The model also struggled with basic instruction-following — told to adopt the name "Jennifer" three times in-conversation, it later denied having that name — and topped out around 2,000-3,000 tokens on a requested 5,000-word essay. Community harnesses have already emerged, including a Rust-based smart-home assistant and OpenBMB's own Electron "desk pet" app with swappable LoRA personalities.

"For a 1B model, this is really amazing that we're starting to get these results where it can consistently act in an intelligent way and be able to do agentic tool calls and reasoning." — Sam Witteveen
Tools: MiniCPM-5, MiniCPM-4.6, Qwen3, GLM-5.2, Artificial Analysis, LoRA
AI Tools
Better Stack

Microsoft Open-Sources SkillOpt to Train Agent Skill Files

Microsoft Research open-sourced SkillOpt, which improves agent performance not by fine-tuning weights or hand-editing prompts, but by treating the agent's markdown skill file itself as a trainable artifact — and it landed best-or-tied-best in all 52 tested settings across seven models, six benchmarks, and two harnesses [7]Better Stack.

Read more

The agent runs a batch of tasks with its current skill file, logging every message, tool call, and score as a "rollout"; a separate optimizer model then reads wins and failures and proposes edits under a strict edit budget (like a learning rate), with no edit accepted unless it beats a held-out validation set. Rejected edits are stored so the optimizer doesn't repeat failed directions. Skills also transfer: a skill.md trained inside Codex, dropped into Claude Code untouched, produced a 31.8-point gain on spreadsheet tasks, and a skill trained on a larger model gave a 15-point boost when moved to a smaller model — evidence the skill captures generalizable strategy rather than model-specific quirks.

"For years we've treated prompts as disposable text you tweak by hand until something works, but skill opt treats them as trainable artifacts and gets results that compete with fine-tuning without ever spinning up a GPU." — Better Stack
Tools: SkillOpt, Codex, Claude Code
Developer Tools
Better Stack

OpenCV 5 Ships a Rewritten DNN Engine — First Major Release Since 2018

OpenCV 5, the first major version bump since 2018, ships a brand-new deep-learning engine that runs YOLO, stable-diffusion-style inpainting, and full vision-language models natively — no PyTorch, no ONNX runtime, nothing bolted on [8]Better Stack.

Read more

The rewrite: 22% to 80% ONNX coverage

The old DNN module processed networks layer by layer and supported only ~22% of ONNX operators; the new engine analyzes the whole network first (shape inference, constant folding, operator fusion), pushing coverage to 80% and letting it handle dynamic shapes and modern transformer architectures the old engine couldn't ~01:02. OpenCV's self-reported CPU benchmarks claim it matches or beats Microsoft's ONNX Runtime — 11.5% faster than YOLOv8, ~37% faster than OWLv2, 30% faster than XFeat ~02:02 — though the host flags these as self-reported and recommends independent benchmarking.

Demos and the catch

Live demos covered zero-dependency colorization, real-time CPU object detection, latent-diffusion inpainting (slow but working), and VLM captioning via PaliGemma — the latter "painfully slow" compared to running a 12B Gemma 4 model on MLX at roughly 100x the speed ~07:04. The new engine is CPU-only at launch, with GPU support planned later in the v5 cycle; developers needing GPU inference today still fall back to the classic CUDA/OpenVINO engine. Verdict: a big leap for running models inside a vision pipeline, but training still belongs to PyTorch.

"OpenCV now has all the pieces you need to run LLMs. The tokenizer, the attention layers, and the KV cache, all built in." — Better Stack
Tools: OpenCV 5, YOLO, ONNX, Microsoft ONNX Runtime, PaliGemma, Gemma 4, MLX, CUDA, OpenVINO
Podcast
AI Engineer

Raphael Kalandadze at AI Engineer: The Missing Layer After Launch

Wandero AI's Raphael Kalandadze argues shipping an agent is now the easy part — the neglected work is operating it in production, and his fix is to have agents operate the agents: a self-built "meta harness" of monitoring, review, and computer-use agents [9]AI Engineer.

Read more

~00:00 Thesis: shipping is where the real work begins. ~02:01 Why agents are hard: endless coverage, losing the feel for the product. ~03:03 Why it's new: non-determinism and hidden failure modes. ~06:05 Core argument: operating an agent is itself an agent problem. ~09:07 Flow 1: a fast log-monitoring agent that auto-opens PRs every 15-60 minutes. ~11:09 A fresh-context review agent scores PRs; the review+PR pipeline ships 10x more PRs than his 3-person team could alone. ~12:09 Flow 2: a session-analyzer health dashboard. ~16:12 A computer-use agent that experiences the product from the user's perspective. ~18:16 The "meta harness" and the closing advice: close the loop first, then remove yourself as the bottleneck.

"Reliable agents shouldn't depend on luck." — Raphael Kalandadze
Tools: Wandero AI
Podcast
AI Engineer

Soheil Feizi at AI Engineer: Verifiable Continual Learning for Agents

RELAI founder Soheil Feizi argues continual learning for agents isn't model fine-tuning — production logs aren't learning environments, and any proposed fix must be lifted into a replayable simulation and proven not to break what already works before it ships [10]AI Engineer.

Read more

~00:01 Thesis: learning without forgetting, across the model, harness, and memory layers. ~02:01 Challenge 1 — where feedback comes from: benchmarks vs. messy production logs. ~04:01 Logs aren't testable on their own; they must be lifted into replayable learning environments. ~06:02 Challenge 2 — acting on feedback across model/harness/memory, with each layer's pitfalls (surveying SFT, DPO, GRPO, RLVR, LoRA, GEPA, Letta, Mem0). ~11:06 Defining "verifiable continual learning" and its four principles: replayability, holisticness, lifelongness, efficiency. ~18:16 Demo: a support-agent benchmark improves from 78% to 97% after a production-log-derived fix. ~21:19 Closing: continual learning isn't fine-tuning, logs aren't environments, and regression-awareness is the frontier.

Tools: RELAI, SFT, DPO, GRPO, RLVR, LoRA, GEPA, Letta, Mem0
Podcast
AI Engineer

Pietro Zullo at AI Engineer: MCP Apps as the New Distribution Surface

Manufact co-founder Pietro Zullo argues that MCP apps — MCP servers returning interactive iframe widgets instead of JSON — plus the newly opened MCP stores are how all software will be built and distributed next: "AI apps are the new browsers, ChatGPTs are the new websites" [11]AI Engineer.

Read more

~00:01 Thesis: MCP apps are how all software will be used. ~03:05 History: from community MCPUI to official MCP apps, and the stores opening. ~06:09 How MCP apps work: sandboxed iframe widgets. ~08:17 Primitives: setState, follow-up messages, token streaming. ~13:20 Redacting data: separating UI output from model output. ~16:24 Live demos across Cursor, Claude/Excalidraw, and ChatGPT. ~21:26 Distribution: self-serve store submission. ~24:28 Dynamic discovery and closing call to ship an MCP app.

Tools: Manufact, MCPUse, Excalidraw MCP, Remotion, Granola, Linear, Cursor, ChatGPT, Claude
Podcast
AI Engineer

Veronica Hylak at AI Engineer: A Three-Part Story Framework for AI Products

Hey AI's Veronica Hylak argues most AI products fail not because the tech is weak but because founders can't explain them, offering a three-step storytelling fix: identify the wound, make the product click, and show the transformation [12]AI Engineer.

Read more

~00:00 Thesis: jargon-heavy pitches kill AI products. ~01:01 Step 1 — identify the wound: lead with the customer's pain, never with what you built. ~02:02 Step 2 — make it click: the "17-year-old test." ~03:02 Viral hooks and picturable metaphors — like "a smoke alarm for AI behavior" — and banning jargon. ~04:03 Step 3 — show the transformation: concrete before/after. ~05:04 Closing: the clearest story wins, demonstrated via an elevator re-pitch.

Tools: Hey AI, Devin
Podcast
Lenny's Podcast

Lenny's Podcast: How Codex Learned to Edit Video

An OpenAI team member recounts an early emergent-use story: their in-house videographer edited the launch videos for the Codex app using Codex itself — which isn't a video editor — after Codex recognized he was using Adobe Premiere Pro and built itself an installable extension to drive it [13]Lenny's Podcast.

Read more

Codex had no video-editing UI, but instead of failing, it built a Premiere Pro extension it could talk to directly — for example instructing it to change a marker inside the app. The speaker cites this as one of the early "whoa, what are people doing with this thing?" moments.

"So, naturally, what Codex then did was built itself an extension that could be installed into Premiere Pro that it then talk to and say, 'Hey, Premiere Pro extension, can you please change this marker inside of the Premiere Pro app?'" — Lenny's Podcast
Tools: Codex, Adobe Premiere Pro
Podcast
Dwarkesh Patel

Dwarkesh Patel x Sarah Paine: Britain's 360-Degree Moat

Historian Sarah Paine argues Britain turned its dependence on trade — a weakness — into strength by using its island geography and a strong navy to avoid a costly standing army, letting it compound wealth while continental rivals bled theirs in constant land wars [14]Dwarkesh Patel.

Read more

As an island, Britain is the only major European power that never needed a large standing army — a competent navy could keep any invading force from ever reaching shore. Trade finances the navy, and the navy protects both homeland and trade in turn. Paine draws a strategic distinction: a navy is a "prevent defeat" strategy, not a "deliver victory" one, since it can't eliminate continental threats outright — but over time, compounding wealth at home versus wealth destroyed abroad in constant warfare tilts the balance decisively in Britain's favor.

"The British are the real pioneers of maritime empire because Britain has this 360 degree moat." — Sarah Paine
"Britain could take a weakness, its dependence on trade, and turn it into strength." — Sarah Paine
Tools: (none — history/geopolitics discussion)
Developer Tools
marimo

marimo: Throwing a GPU at the Traveling Salesman Problem

A marimo tutorial reframes the classic traveling-salesman problem as a matrix of swap savings so it can be solved with massively parallel GPU compute, using a nearest-neighbor windowing trick to keep the matrix in memory regardless of instance size [15]marimo.

Read more

Instead of treating a tour as a graph, every possible city-pair "2-opt" swap's savings can be represented as one symmetric matrix — a reframing that unlocks GPU parallelism, since each matrix entry can be computed independently ~01:01. To keep the (otherwise impractically large) matrix in GPU memory, each city's candidates are restricted to a fixed window of nearest neighbors — a tunable knob that also extends to pricier 3-opt (three-edge) swaps ~02:02. In a live demo, GPU 2-opt landed about 14% off the known optimum in roughly half a second — and surprisingly, a purely random starting tour worked nearly as well as a "clever" heuristic start ~05:04. Switching to 3-opt hit the optimal value exactly on that instance. Against Google's OR-Tools, both approaches matched on moderate problems, though the GPU heuristic landed ~12% off optimal on a very large benchmark instance — not quite apples-to-apples, since OR-Tools solves the broader vehicle-routing problem. The host recommends JAX for this style of GPU programming, citing its JIT compiler and vmap auto-vectorization, and points to Molab for a free GPU to experiment on.

"Even though we have a pretty dumb algorithm, we are able to get away with it simply because we throw a GPU at the problem." — marimo
Tools: marimo, JAX, OR-Tools, Molab
Developer Tools
Real Python

Declarative vs. Imperative Prompting for AI Agents

Real Python argues against boxing agents in with rigid, step-by-step instructions — the better skill/system-prompt design tells an agent what tools exist and when they're relevant, then lets it determine its own sequence of steps [16]Real Python.

Read more

The speakers describe having to "unlearn" imperative-style prompting ("do this, then this, then this"), which doesn't generalize across different agents and situations. They compare the shift to mentoring a junior engineer over time: initially you supervise closely with explicit instructions, but as trust builds you shift to pointing them at resources and letting them exercise judgment.

"Don't try to put your agent in a box." — Real Python
Tools: (none named)
Developer Tools
Arjay McCandless

Why Are Databases So Fast? B-Trees and B+ Trees Explained

A short explainer on why databases outperform naive binary search trees at lookups: B-trees hold multiple sorted values per node to keep the tree shallow, and B+ trees link leaf nodes in sorted order to make range queries fast [17]Arjay McCandless.

Read more

Unlike a binary search tree where each node has only two children, a B-tree node holds multiple sorted values and points to many children, shortening the tree and reducing traversals. For range queries, B+ trees store all actual values only in leaf nodes and link those leaves together in sorted order, so a range query starts at the first matching leaf and simply follows pointers sequentially rather than repeatedly climbing back up the tree.

Tools: (none — conceptual database internals)
Developer Tools
The Pragmatic Engineer

Why Kent Beck Named It "Extreme Programming"

Kent Beck explains the name by analogy to extreme sports: you don't hop on a snowboard down an avalanche on your first try — you must be supremely prepared, and that preparation is the "extreme" part [18]The Pragmatic Engineer.

Read more

For the "programming" half, Beck says he wanted to describe the moment of sitting with fingers on the keyboard staring at code — the point where you can no longer fool yourself into thinking you understand something, since the program either computes the correct value or it doesn't.

"You don't just hop on a snowboard at the top of some avalanche the first time. You have to be supremely prepared. So, that's the extreme part." — Kent Beck
Tools: (none — software history anecdote)
AI Tools
Github Awesome

$40K Home Rig: Four RTX Pro 6000s to Self-Host Opus-Class Inference

An engineer documents a roughly $40,000 build log using four RTX Pro 6000 GPUs linked via PCIe switches, aimed at running Claude Opus-class inference entirely on owned hardware — no subscription, no API key [19]Github Awesome.

Read more

The four cards communicate directly with each other rather than routing through the CPU. The write-up walks through the specific BIOS settings, kernel flags, and an ACS (Access Control Services) fix required to unlock full card-to-card bandwidth — without it, GPU-to-GPU communication was effectively broken; applying it restored full Gen 4 PCIe speeds.

"$40,000 gets you noticeably closer to Claude Opus running entirely on your own hardware. No subscription, no API key." — Github Awesome
Tools: RTX Pro 6000, Claude Opus
AI Models
AI Search

Google Ships Efficiency-Focused Nano Banana 2 Light and Gemini Omni Flash

While other labs race for raw power, Google shipped two efficiency-first models: Nano Banana 2 Light generates an image in 4 seconds for ~3 cents per 1K images (roughly half the price of Nano Banana 2), and Gemini Omni Flash edits video with natural language [1]AI Search.

Read more

Gemini Omni Flash accepts existing videos as references and edits with natural language — adding, replacing objects, changing style, adding VFX. Google's self-reported benchmarks claim it beats Alibaba's Happy Horse, Cling, and Seed Dance, but the presenter calls those numbers "misleading" based on his own testing, where Seed Dance looked clearly better than the other two. Both models are free to try on Google Flow with daily-resetting limits.

Tools: Nano Banana 2 Light, Gemini Omni Flash, Google Flow
AI Future
AI Search

Ubtech's U1: $150K Ultra-Realistic Companion Robots, 13,000+ Pre-Orders

Ubtech unveiled the U1 line of ultra-realistic humanoid companion robots — visible pores, blood vessels, and fingerprints — with "emotional AI" recognizing 20+ human emotions at over 90% accuracy, priced from ~$18,000 to ~$150,000 for a full-body version [1]AI Search.

Read more

The line has already logged over 13,000 pre-orders — more than the Unitree G1 robot received — with deliveries planned within the year. The presenter jokes about the relationship implications while also reflecting seriously on how ultra-realistic companion robots could reshape human-to-human and human-to-robot relationships.

Tools: Ubtech U1, Unitree G1
Developer Tools
AI Search

ComfyUI Ships Comfy MCP to Let Agents Drive Its Node Editor

ComfyUI released Comfy MCP, a bridge that lets AI agents operate its notoriously fiddly node-based interface via natural language — searching models, nodes, and templates, then building and running entire workflows autonomously [1]AI Search.

Read more

Rather than manually building nodes and hunting for models, an agent can search and run entire workflows from a chatbot-style prompt, with Comfy auto-updating best-practice workflows so the agent stays current. It's currently in free public beta with docs for various agent platforms.

Tools: Comfy MCP, ComfyUI
AI Future
AI Search

Meta's Brain2Qwerty Hits 78% Word Accuracy Reading Thoughts via MEG

Meta's non-invasive brain-to-text project decodes MEG (magnetoencephalography) brain signals into typed text with no surgery or implants, reaching roughly 69% character accuracy and 78% word accuracy in its second version [1]AI Search.

Read more

Version two was trained on over 10x more data than version one, substantially improving accuracy. The code is already released with install instructions.

Tools: Brain2Qwerty
AI Models
Hugging Face Daily Papers

arXiv's Agentic-Evaluation Glut: Six New Benchmarks in One Day

Six separate papers landed on Hugging Face's Daily Papers the same day, each proposing new benchmarks or testbeds for evaluating agentic AI — a sign research effort is shifting from building agents toward measuring whether they actually work over long horizons [20]Hugging Face Daily Papers.

Read more

The cluster: AgenticSTS (a bounded-memory testbed for long-horizon agents), EvoPolicyGym (evaluating autonomous policy evolution), AgenticDataBench (a benchmark for data agents), SkillCoach (self-evolving rubrics for agentic skill-use), AutoMem (automated learning of memory as a cognitive skill), and PACE (a proxy for agentic capability evaluation). None share a single lab, but they converge on the same gap: as agents take on longer-horizon, memory-dependent tasks, single-turn benchmarks miss failure modes like memory drift, skill regression, and policy evolution over time. Also notable from the same day: WorldDirector, a world simulator with persistent dynamic memory, and Multi-Resolution Flow Matching, a training-free diffusion acceleration technique.

Tools: AgenticSTS, EvoPolicyGym, AgenticDataBench, SkillCoach, AutoMem, PACE, WorldDirector
AI Tools
AI Search

AI Search Roundup: One-Step Image Gen, Robot Dexterity, and More

The rest of the day's AI Search roundup covered a dense slate of smaller releases spanning image generation, robotics, and simulation [1]AI Search.

Read more
  • MuseViT — a <500MB open model purpose-built to read sheet music, trained on 9.7M pages via masked reconstruction.
  • LiveEdit — real-time video editing at ~13fps from a reference video plus text prompt, via a causal chunk-wise streaming architecture; 17GB, code released.
  • VidiHand — 3D hand reconstruction from video that stays accurate even with partial occlusion, useful for robotics training data and AR/VR.
  • OmniContact — a "contact flow" framework letting humanoid robots chain locomotion and manipulation over ~40-minute horizons with near-100% box-carrying success.
  • PhysFormer — predicts physically realistic 3D object motion (rigid vs. elastic materials) from a starting mesh, trained on 100,000+ simulated trajectories.
  • Nvidia Aspire — a self-improving robotics system that writes robot control code, repairs it on failure, and saves corrected behaviors to a growing skill library.
  • RDM (Representation Distribution Matching) — generates images in a single step instead of the usual 30-50, by matching high-level visual feature distributions.
  • Mr. Flow — training-free, model-agnostic speedup (up to 21x) by generating low-res then spending one high-res cleanup step.
  • SimFoundry — turns one photo or video into a physically accurate, simulation-ready 3D scene for robot training, with scene variation generation for generalization.
  • Cord — teaches robots dexterous hand skills from human demos by representing them as forces/torques on objects, transferable across differently-built robot hands.
  • Luna (Meta) — animates 3D human avatars from a few images plus any driving signal (video, skeleton, or line art) using Gaussian deformations instead of a pose skeleton.
Tools: MuseViT, LiveEdit, VidiHand, OmniContact, PhysFormer, Nvidia Aspire, RDM, Mr. Flow, SimFoundry, Cord, Luna
Industry
EO EO

EO: Two Founder Origin Stories

Two founder-interview clips from EO: one describes building a company from a college library booth at night while stuck in Australia awaiting a US visa [21]EO, the other recounts how a 911-dispatch live-streaming product saved a life in its very first week live [22]EO.

Read more

Building from a library booth

The founder describes having no office in the early days, working from a booth in the library of the college he'd dropped out of, waking at 11pm to overlap his schedule with San Francisco business hours while waiting on a US visa from Australia. The clip doesn't name the company or explain how it reached the $65M figure referenced in its title.

Saving a life in week one

A different founder recounts turning on their product for 911 dispatch centers and, within the first week, a bystander used its live-video-streaming feature so a call taker could watch and coach CPR in real time on someone found unconscious — the person survived. The company is now on track to process over 30 million calls this year.

"I remember when we first turned on our product for 911 centers. Within our first week, people used it to save a life." — EO
Tools: (none named)
Industry
Morning Brew

America's 250th: Iconic Brands, the Next 250 Years, and the Made-in-USA Myth

A three-part Morning Brew package for July 4th's 250th anniversary looks back at the brands that defined American business, speculates on the companies that could define the next 250 years, and explains why "Made in the USA" labeling is so rare [23]Morning Brew.

Read more

The last 250 years

From Coca-Cola (~40% of the nonalcoholic beverage market, founded by "a morphine-addicted chemist and former Confederate soldier") to Ford and GM's assembly lines — including the 1936 Flint sit-down strike where 136,000 workers over 44 days won UAW recognition — to General Electric's rise from Edison's legacy and Big Tech's trillion-dollar milestones (Apple hit $1T in 2018), the piece rounds up brands "that will flash before our eyes on our deathbeds."

The next 250 years

Predicting out 250 years is nearly impossible (Nvidia's own rapid rise is the cautionary example), but likely contenders include OpenAI, Anthropic, and Amazon's cloud arm in AI; Alphabet, called "especially well positioned" via Gemini, stakes in Anthropic and SpaceX, quantum computing, and Waymo [24]Morning Brew; energy incumbents Exxon and Chevron facing pressure from fusion startups amid data-center demand growth; and Johnson & Johnson in healthcare, on the logic that "death will probably still be around in 250 years."

The Made-in-USA myth

The FTC requires "Made in the USA" products to be "all or virtually all" domestic, but 14 US free-trade agreements with 20 countries and the fact that 11% of US export value derives from foreign inputs push companies like Levi's, Rawlings, and Harley-Davidson toward cheaper foreign labor [25]Morning Brew. The piece notes 108,000 US manufacturing jobs were eliminated in the first year of Trump's second term, framing his 2025 tariff push as an uncertain attempt to reverse the trend.

"The only thing more American than jeans, baseballs, and motorcycles is maximizing profits and undercutting competitors by using cheaper foreign labor." — Morning Brew
Tools: (none — business/culture feature)

Sources

  1. YouTube Full body waifus, Claude Fable is back, LongCat 2.0, mind-reading AI, live video editing: AI NEWS — AI Search, Jul 5
  2. Blog sqlite-utils 4.0rc2, mostly written by Claude Fable — Simon Willison, Jul 5
  3. YouTube Caveman + Fable 5: This SIMPLE Trick makes Fable cheaper than Opus! — AICodeKing, Jul 5
  4. YouTube You Can't Compete on Cheap Models Anymore — AI News & Strategy Daily | Nate B Jones, Jul 5
  5. YouTube Fable 5 doesn't want your prompt. It wants the whole job. — AI News & Strategy Daily | Nate B Jones, Jul 5
  6. YouTube MiniCPM5 - The 1B Cognitive Core? — Sam Witteveen, Jul 5
  7. YouTube Microsoft Just Open-Sourced a Cheat Code for AI Agents (SkillOpt) — Better Stack, Jul 5
  8. YouTube OpenCV 5 Is Here - The Biggest Update Since 2018 (We Tested It) — Better Stack, Jul 5
  9. YouTube The Missing Layer After Launch - Raphael Kalandadze, Wandero AI — AI Engineer, Jul 5
  10. YouTube Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI — AI Engineer, Jul 5
  11. YouTube MCP Apps: Primitives, discovery, and the Future of Software - Pietro Zullo, Manufact, Inc — AI Engineer, Jul 5
  12. YouTube Your AI Product Will Fail Unless You Can Explain It - Veronica Hylak, Hey AI — AI Engineer, Jul 5
  13. YouTube How Codex learnt to edit videos — Lenny's Podcast, Jul 5
  14. YouTube Britain turned its biggest weakness into the source of its power - Sarah Paine — Dwarkesh Patel, Jul 5
  15. YouTube How to Throw a GPU at a Problem — marimo, Jul 5
  16. YouTube Declarative vs Imperative Prompting for AI Agents — Real Python, Jul 5
  17. YouTube Why are databases so fast? — Arjay McCandless, Jul 5
  18. YouTube Why Kent Beck named it "Extreme Programming" — The Pragmatic Engineer, Jul 5
  19. YouTube $40k home Opus: one engineer's full build log for four RTX Pro 6000s — Github Awesome, Jul 5
  20. Blog Daily Papers — Hugging Face, Jul 5
  21. YouTube He worked nights in the library and built a $65M Company — EO, Jul 5
  22. YouTube 25-year-old who sold for $640M: In week one, it saved a life — EO, Jul 5
  23. Newsletter The most iconic US brands of the last 250 years — Morning Brew, Jul 5
  24. Newsletter US companies that will define the next 250 years — Morning Brew, Jul 5
  25. Newsletter The state of 'Made in the USA' products — Morning Brew, Jul 5