> ## Content Index
> Fetch the complete content index at: https://genaisecretsauce.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# GenAI Secret Sauce Daily Digest - 2026-06-24
- URL: https://genaisecretsauce.com/genai-secret-sauce-daily-digest-2026-06-24/
- Published: 2026-06-24T23:22:39.000Z
- Updated: 2026-07-12T21:04:36.000Z
- Description: OpenAI Builds Its First Chip - and Microsoft Wants 40% of Production · Google Puts Screen Control Directly Into Its Cheapest AI Model · The NSA Lost Access to America's Most Capable AI - Because of Its Own Government's Export Controls
- Author: Jasmine Robinson
- Tags: Daily Digest

Watch today's digest as a video summary (generated by NotebookLM)

By the Numbers

## Statistically Speaking

[40%](https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom?ref=genaisecretsauce.com) [of initial production](https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom?ref=genaisecretsauce.com) 

OpenAI Builds Its First Chip - and Microsoft Wants 40% of Pr

Top Story

[3](https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom?ref=genaisecretsauce.com) [years to under a year](https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom?ref=genaisecretsauce.com) 

OpenAI Builds Its First Chip - and Microsoft Wants 40% of Pr

[88%](https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-powerful-mythos-ai-reportedly-breached-almost-all-nsa-classified-systems-within-a-few-hours-during-red-team-test-report-sheds-more-light-on-the-u-s-governments-sudden-ban-on-the-flagship-models?ref=genaisecretsauce.com) [odds of Fable 5 returning by July](https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-powerful-mythos-ai-reportedly-breached-almost-all-nsa-classified-systems-within-a-few-hours-during-red-team-test-report-sheds-more-light-on-the-u-s-governments-sudden-ban-on-the-flagship-models?ref=genaisecretsauce.com) 

The NSA Lost Access to America's Most Capable AI - Because o

[2](https://www.greptile.com/blog/prs-on-openclaw?ref=genaisecretsauce.com) [to 3,400 between December and February](https://www.greptile.com/blog/prs-on-openclaw?ref=genaisecretsauce.com) 

AI Agent Pull Requests Now Look Like Email Spam in 2000

[90%](https://www.greptile.com/blog/prs-on-openclaw?ref=genaisecretsauce.com) [of AI](https://www.greptile.com/blog/prs-on-openclaw?ref=genaisecretsauce.com) 

AI Agent Pull Requests Now Look Like Email Spam in 2000

[$175](https://www.latent.space/p/databricks?ref=genaisecretsauce.com) [billion valuation target reflects Databricks' bet that](https://www.latent.space/p/databricks?ref=genaisecretsauce.com) 

Databricks Open-Sources an Operating System for Enterprise A

One Thing to Tell Your Friends

## One Thing to Tell Your Friends

OpenAI just built its own computer chip in nine months - and Microsoft is buying almost half of them.

Summary

## TL;DR

Top Stories

[OpenAI Builds Its First Chip](https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom?ref=genaisecretsauce.com), [Google Puts Screen Control Directly Into Its Cheapest AI Model](https://deepmind.google/blog/introducing-computer-use-in-gemini-3-5-flash?ref=genaisecretsauce.com), and [The NSA Lost Access to America's Most Capable AI](https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-powerful-mythos-ai-reportedly-breached-almost-all-nsa-classified-systems-within-a-few-hours-during-red-team-test-report-sheds-more-light-on-the-u-s-governments-sudden-ban-on-the-flagship-models?ref=genaisecretsauce.com).

Trends

**The Custom Silicon Arms Race Is Accelerating**, **Computer Use Is Becoming a Commodity Feature**, and **AI Safety Research Is Finding Dangerous Blind Spots**.

Creative AI

[Krea 2 Turbo: Open](https://huggingface.co/krea/Krea-2-Turbo?ref=genaisecretsauce.com) and [DiffusionGemma: Google's Multimodal Generation Model](https://huggingface.co/google/diffusiongemma-26B-A4B-it?ref=genaisecretsauce.com).

Dev Tools

[RubyLLM: One Interface for 800+ AI Models](https://rubyllm.com/?ref=genaisecretsauce.com), [NVIDIA NeMo AutoModel: 3.69x Faster Fine](https://huggingface.co/blog/nvidia/accelerating-fine-tuning-nvidia-nemo-automodel?ref=genaisecretsauce.com), and [JupOtter: Bug Detection Built for Jupyter Notebooks](https://arxiv.org/abs/2606.23877?ref=genaisecretsauce.com).

Research

[A Drop-In Fix Cuts Long](https://arxiv.org/abs/2606.24467?ref=genaisecretsauce.com), [Small AI Models Can Match Giants at Reading](https://arxiv.org/abs/2606.23695?ref=genaisecretsauce.com), and [OpenThoughts](https://arxiv.org/abs/2606.24855?ref=genaisecretsauce.com).

Business

[OpenAI Enters the Chip Business](https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom?ref=genaisecretsauce.com), [Databricks Targets $175 Billion Valuation](https://www.latent.space/p/databricks?ref=genaisecretsauce.com), and **AI Economists Are Here: Labs Hiring Philosophers in Droves**.

Education

[The Mythos/Fable Crisis Reshapes AI Governance Debates in Academia](https://aiandacademia.substack.com/p/on-the-epic-and-still-developing), [AI](https://simonwillison.net/2026/Jun/24/tom-macwright/?ref=genaisecretsauce.com#atom-everything), and [LeetCode Persists Because It Tests Scalability, Not Prediction](https://newsletter.pragmaticengineer.com/p/tech-interviews-with-neetcode?ref=genaisecretsauce.com).

Surprising

[AI Models Drop Their Scientific Rigor the Moment You Ask for Advice](https://arxiv.org/abs/2606.24370?ref=genaisecretsauce.com), [LLM Mental Health Safeguards Fail Completely for Eating Disorders](https://arxiv.org/abs/2606.23884?ref=genaisecretsauce.com), and [Physicists Say LLM Scaling Is Thermodynamically Unsustainable](https://arxiv.org/abs/2606.24504?ref=genaisecretsauce.com).

Worth Watching

[Cryptographic Proof Could Make AI Agents Auditable by Default](https://arxiv.org/abs/2606.23768?ref=genaisecretsauce.com), [FP8 Tensor Core Tricks Could Unlock the Next Generation of GPUs](https://arxiv.org/abs/2606.23698?ref=genaisecretsauce.com), and [Bayesian Control Could Make Coding Agents Dramatically Cheaper](https://arxiv.org/abs/2606.24453?ref=genaisecretsauce.com).

GitHub

Leading repos: [calesthio/OpenMontage](https://github.com/calesthio/OpenMontage?ref=genaisecretsauce.com) (+3,703), [ZhuLinsen/daily\_stock\_analysis](https://github.com/ZhuLinsen/daily%5Fstock%5Fanalysis?ref=genaisecretsauce.com) (+1,461), and [NousResearch/hermes](https://github.com/NousResearch/hermes-agent?ref=genaisecretsauce.com) (+1,174).

HuggingFace

Leading models: [deepseek-ai/DeepSeek-V4](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro?ref=genaisecretsauce.com) (2.05M), [zai-org/GLM](https://huggingface.co/zai-org/GLM-5.2?ref=genaisecretsauce.com) (57.2K), and [MiniMaxAI/MiniMax](https://huggingface.co/MiniMaxAI/MiniMax-M3?ref=genaisecretsauce.com) (143K).

Product Hunt

Top launches: [Propane](https://www.producthunt.com/products/propane?ref=genaisecretsauce.com) (437), [Tencent EdgeOne Makers](https://www.producthunt.com/products/tencent-edgeone-makers?ref=genaisecretsauce.com) (357), and [Buy by Agentcard](https://www.producthunt.com/products/buy-by-agentcard?ref=genaisecretsauce.com) (154).

API Pricing

**What this means:** The price floor continues to fall.

arXiv

[CompressKV: Semantic-Retrieval-Guided KV-Cache Compression for Resource-Efficient Long](https://arxiv.org/abs/2606.24467?ref=genaisecretsauce.com) — 97% of full-cache performance at 3% memory cost on LongBench QA.

FYI

## Hot off the Presses

01

### OpenAI Builds Its First Chip - and Microsoft Wants 40% of Production

**What this means for you:** The companies behind ChatGPT are now building their own silicon, which could eventually make AI services cheaper and faster as they stop paying NVIDIA's markup.

OpenAI and Broadcom jointly announced Jalapeño - an "Intelligence Processor" designed from scratch for running AI models (inference), not training them. The nine-month design-to-tape-out timeline is believed to be the fastest ASIC development cycle ever achieved for a chip of this complexity, accelerated by using AI tools in the design process itself.

The move follows Google (TPU), Amazon (Trainium/Inferentia), and Meta (MTIA) in the trend of AI companies building custom hardware. OpenAI was the last major frontier lab relying entirely on NVIDIA GPUs.

- **Microsoft is purchasing approximately 40% of initial production** \- a massive endorsement from OpenAI's largest partner and investor
- **Gigawatt-scale deployment is targeted for late 2026** \- meaning data centers consuming as much power as a mid-sized city
- **The chip optimizes for inference, not training** \- reflecting the industry shift as serving billions of users becomes the dominant cost
- **AI-assisted chip design** cut the development cycle from the typical 2-3 years to under a year

[Techcrunch →](https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom?ref=genaisecretsauce.com)

02

### Google Puts Screen Control Directly Into Its Cheapest AI Model

**What this means for you:** An AI that can click buttons, fill forms, and navigate apps on your behalf is now built into Google's fastest and most affordable model - not a premium add-on.

Google DeepMind merged computer use as a native capability into Gemini 3.5 Flash, their budget-tier model optimized for speed. Previously, computer use required a separate, dedicated Gemini 2.5 model. Now browser automation, mobile app control, and desktop navigation are available from the same model used for chat and coding.

The significance is the price tier: putting computer use in Flash rather than Pro signals that Google sees screen control as a commodity feature, not a premium one.

- **Native integration** means no switching between models for different tasks - one Application Programming Interface (API) call handles both text analysis and screen interaction
- **Adversarial injection training** teaches the model to resist prompt injection attacks through web pages it navigates - a safety measure competitors haven't publicly matched
- **Automatic task-halt safety measures** stop the agent if it detects it's being manipulated
- **Direct competition** with Anthropic's computer use (available since late 2024) and OpenAI's Operator

[Google DeepMind →](https://deepmind.google/blog/introducing-computer-use-in-gemini-3-5-flash?ref=genaisecretsauce.com)

03

### The NSA Lost Access to America's Most Capable AI - Because of Its Own Government's Export Controls

**What this means for you:** A rule designed to keep powerful AI away from foreign adversaries accidentally cut off U.S. intelligence analysts from the tools they depend on.

*Previously: [June 23](https://genaisecretsauce.com/genai-secret-sauce-daily-digest-2026-06-23/) \- Fable 5 restrictions entered their second week after the Commerce Department barred foreign nationals from accessing Anthropic's Mythos and Fable models.*

**Today:** The New York Times reported that parts of the NSA have lost access to Anthropic's Mythos 5, the model that - during a controlled red-team exercise - breached "almost all" of the agency's classified systems "not in weeks, but in hours." The irony is acute: the same government that witnessed the model's power firsthand is the one whose export controls forced Anthropic to pull it globally.

- **Anthropic couldn't enforce nationality-based access restrictions** without pulling the models for everyone, including U.S. government users
- **Prediction markets give 88% odds** of Fable 5 returning by July 31, according to Zvi Mowshowitz's analysis
- **The agency may retain access to older model versions** but loses updates, support, and the most capable models
- **Multiple analysts describe this as "a train wreck"** of policy implementation

[Tomshardware →](https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-powerful-mythos-ai-reportedly-breached-almost-all-nsa-classified-systems-within-a-few-hours-during-red-team-test-report-sheds-more-light-on-the-u-s-governments-sudden-ban-on-the-flagship-models?ref=genaisecretsauce.com)[Analysis →](https://thezvi.substack.com/p/the-once-and-future-fable-4)

04

### AI Agent Pull Requests Now Look Like Email Spam in 2000

**What this means for you:** Open-source projects are being flooded with low-quality contributions from AI coding agents, and maintainers don't yet have the tools to filter them.

Greptile analyzed pull request patterns in the OpenClaw repository - which became the fastest-growing GitHub repo in history - and found a pattern that mirrors the early internet's spam crisis.

The parallel to email spam is instructive: the technology that solved email spam (Bayesian filters, reputation systems, rate limiting) took years to develop. Open-source is now facing the same reckoning, but with contributions that look superficially legitimate.

“The spam filter hasn't been invented yet.”

- **Weekly PR volume exploded from \~2 to 3,400** between December and February - a 1,700x increase
- **Merge rate collapsed to 9.3%** \- meaning over 90% of AI-generated contributions were rejected
- **Multiple AI agents independently submitted identical PRs** for the same issues, with no coordination
- **Maintainer burden grew faster than the project** \- review time per PR didn't decrease, but volume made the queue unmanageable

[Greptile →](https://www.greptile.com/blog/prs-on-openclaw?ref=genaisecretsauce.com)

05

### Databricks Open-Sources an Operating System for Enterprise AI Agents

**What this means for you:** If your company uses AI coding tools from different vendors, Databricks just released a free system that lets all of them work together through one interface.

Databricks cofounders Matei Zaharia and Reynold Xin, speaking on the Latent Space podcast, unveiled Omnigent - an open-source meta-harness that lets agents from Claude Code, Codex, Cursor, and other systems work through unified APIs and session management. The company received over 400 pull requests within days of launch.

- **Universal agent compatibility** \- one harness for agents from any vendor, avoiding lock-in
- **Session management and audit trails** built in for enterprise compliance
- **$175 billion valuation target** reflects Databricks' bet that the agent infrastructure layer is as valuable as the cloud infrastructure layer
- **Enterprise-grade security controls** including access policies and data governance

[Latent Space →](https://www.latent.space/p/databricks?ref=genaisecretsauce.com)

Trends & Themes

## Trends & Themes

[![Trends & Themes](https://genaisecretsauce.com/content/images/2026/06/section-what-this-means-2026-06-24-1.png)](https://genaisecretsauce.com/content/images/2026/06/section-what-this-means-2026-06-24-1.png) 

### The Custom Silicon Arms Race Is Accelerating

**Why this matters to you:** As AI companies build their own chips, the cost of using AI services should fall - and the companies that get hardware right will have a permanent advantage.

The pattern is clear: training still needs NVIDIA's most powerful GPUs, but inference - the part that serves billions of users - is being commoditized through custom silicon. NVIDIA's inference monopoly is eroding from multiple directions simultaneously.

- **OpenAI's Jalapeño is the fifth major AI company** to announce custom inference silicon, following Google (TPU), Amazon (Trainium), Meta (MTIA), and Microsoft (Maia)
- **Microsoft buying 40% of Jalapeño production** suggests the economics already beat NVIDIA for inference at scale
- **FP8 tensor core innovations** (a new paper shows how to get FP64-equivalent precision from FP8 hardware) could make next-generation chips even more efficient

### Computer Use Is Becoming a Commodity Feature

**Why this matters to you:** The ability to have an AI control your screen, click buttons, and fill forms is moving from experimental to standard - and it's getting cheaper fast.

Computer use following the same path as code generation: first a novelty, then a premium feature, then table stakes.

- **Google added computer use to Flash** (their cheapest model), signaling it's no longer a premium capability
- **A new benchmark paper (GUI vs. CLI)** found CLI agents hit 69.3% success on desktop tasks vs. 59.3% for GUI agents - but the gap is narrowing and fixable through better skill coverage
- **RL training for computer-use agents** showed a 12.6 percentage point improvement using autonomous vision-language evaluation - no human labeling needed
- **Three competing approaches** are now live: Anthropic (separate API), Google (built into Flash), and OpenAI (Operator as a product)

### AI Safety Research Is Finding Dangerous Blind Spots

**Why this matters to you:** Researchers are discovering that AI models behave safely in lab tests but fail in ways that matter when people actually use them for real decisions.

The pattern: safety evaluations that test models in controlled settings consistently overstate real-world safety. The gap between lab and deployment is a measurement failure, not a model failure.

- **LLMs maintain causal caution 91-100% of the time in academic contexts** but only 0.5-18% when users ask for practical advice - a one-line self-correction prompt restores it to 71-100%
- **Large Language Model (LLM) mental health safeguards hold for suicide** but fail at rates up to 100% for eating disorders and substance use under adversarial prompting
- **Self-recognition finetuning** can both prevent and reverse "emergent misalignment" - the phenomenon where fine-tuning on benign data causes harmful behavior - by stabilizing the model's identity

### Agent Infrastructure Is Becoming Its Own Software Category

**Why this matters to you:** Just as cloud computing created a new layer of infrastructure companies (AWS, Docker, Kubernetes), AI agents are spawning their own infrastructure ecosystem.

When debugging, memory management, and orchestration all have dedicated research papers and open-source tools, you're looking at a new software category forming.

- **Databricks Omnigent** provides universal agent orchestration for enterprises
- **A new paper (MemClaw)** formalizes four failure modes when multiple agents share memory: leakage, staleness, contradiction, and lost provenance
- **SAFARI** solves agent debugging at scale - attributing failures in million-token execution traces with 20% better accuracy than existing methods
- **Bayesian control for coding agents** reframes agent orchestration as cost-sensitive hypothesis testing, improving cost-performance tradeoffs

### Scaling Efficiency Now Matters More Than Scaling Size

**Why this matters to you:** The era of "just make the model bigger" is giving way to an era of "make the same model cheaper to run" - which should eventually lower prices for everyone.

- **CompressKV retains 97% of model accuracy using only 3% of the KV-cache** \- a 30x memory reduction for long-context inference with no retraining
- **Plasticity loss follows a sublinear scaling law** \- bigger models delay but never prevent the inability to learn new information, challenging the assumption that scale solves everything
- **Task-specific distillation** shows general benchmarks collapse before domain benchmarks under pruning, meaning small specialized models can outperform large general ones
- **A physics-informed analysis** argues LLM scaling exponents are "too small to be sustainable" from an energy standpoint

Creative AI & Media

## Creative AI & Media

### Krea 2 Turbo: Open-Weight Image Generation in Under 2 Seconds

**What this means for you:** A new open-source image generator matches commercial quality while running fast enough for real-time creative workflows.

- **12 billion parameters** generating up to 2048x2048 images
- **Sub-2-second generation** on consumer GPUs
- **Open weights** allow local deployment without API costs

[HuggingFace →](https://huggingface.co/krea/Krea-2-Turbo?ref=genaisecretsauce.com)

### DiffusionGemma: Google's Multimodal Generation Model

**What this means for you:** Google released an open model that generates text and images together, with only 3.8 billion active parameters.

- **26B total, 3.8B active parameters** (Mixture-of-Experts design)
- **Generates 1,100+ tokens** worth of multimodal content in a single pass
- **Discrete diffusion architecture** \- a departure from autoregressive generation

[HuggingFace →](https://huggingface.co/google/diffusiongemma-26B-A4B-it?ref=genaisecretsauce.com)

Developer Tools

## Developer Tools & Infrastructure

### RubyLLM: One Interface for 800+ AI Models

**What this means for you:** Ruby developers now have a mature, minimal framework that works with every major AI provider through a single API.

- **13+ providers** including OpenAI, Anthropic, Google, AWS Bedrock, DeepSeek, Mistral, and Ollama
- **Only 3 runtime dependencies** \- deliberately minimal footprint
- **v1.16.0** with 324 stars on Hacker News today

[Try it →](https://rubyllm.com/?ref=genaisecretsauce.com)[Rubyllm →](https://rubyllm.com/?ref=genaisecretsauce.com)

### NVIDIA NeMo AutoModel: 3.69x Faster Fine-Tuning With One Import

**What this means for you:** Fine-tuning large AI models just got dramatically cheaper - NVIDIA's new library cuts training time by nearly 4x on existing hardware.

- **3.69x speedup and 29% memory reduction** for Qwen3-30B fine-tuning on 8x H100 GPUs
- **Expert Parallelism + DeepEP** optimizations for Mixture-of-Experts models
- **Drop-in replacement** for HuggingFace Transformers - change one import line

[Hugging Face: blog/nvidia →](https://huggingface.co/blog/nvidia/accelerating-fine-tuning-nvidia-nemo-automodel?ref=genaisecretsauce.com)

### JupOtter: Bug Detection Built for Jupyter Notebooks

**What this means for you:** The most popular tool for data science just got a dedicated bug finder that understands notebook-specific failure modes.

- **Cell-aware tokenization** catches bugs that span multiple notebook cells
- **Beats both static analyzers and LLMs** on 2 of 3 benchmarks
- **21,000-notebook labeled dataset** released alongside the tool

[arXiv →](https://arxiv.org/abs/2606.23877?ref=genaisecretsauce.com)

### VeriPilot: LLM-Powered Hardware Debugging

**What this means for you:** Chip designers can now use AI to debug hardware designs, with a 31-point accuracy improvement over raw GPT-4o.

- **85.71% debugging accuracy** (up from GPT-4o's 54.3%) by injecting structured circuit analysis
- **Handles Verilog** \- the dominant language for chip design
- **Traces bugs through complex signal dependencies** that stump standard AI approaches

[arXiv →](https://arxiv.org/abs/2606.23759?ref=genaisecretsauce.com)

Research & Models

## Research & Models

### A Drop-In Fix Cuts Long-Context AI Memory by 97% With Almost No Accuracy Loss

**What this means for you:** Running AI on long documents is about to get dramatically cheaper - this technique needs no retraining and works on existing models.

CompressKV identifies "Semantic Retrieval Heads" - the specific parts of an AI model that actually find important information - and focuses all memory on those, discarding the rest.

“97% accuracy at 3% memory cost”

- **97% of full performance retained using only 3% of memory** on long-document question answering
- **90% accuracy with just 0.7% memory** on needle-in-a-haystack tests
- **No retraining required** \- works as a drop-in layer for existing deployments

[arXiv →](https://arxiv.org/abs/2606.24467?ref=genaisecretsauce.com)

### Small AI Models Can Match Giants at Reading - They Just Ignore What You Give Them

**What this means for you:** When you ask an AI to answer based on documents you provide, larger models are more likely to ignore your documents and answer from memory instead.

- **1.5 billion parameter models can match 72 billion** at factual extraction from provided documents
- **Larger models override provided evidence with stored knowledge** in approximately 50% of adversarial tests
- **A new metric (NCU)** reveals whether AI is actually reading your documents or reciting memorized answers

[arXiv →](https://arxiv.org/abs/2606.23695?ref=genaisecretsauce.com)

### OpenThoughts-Agent: A Fully Open Recipe for Training AI Agents

**What this means for you:** The recipe for building capable AI agents is now fully public - models, data, training code, and 100+ experiments.

- **Qwen3-32B fine-tuned to 44.8% average accuracy** across 7 agent benchmarks - 3.9 points above the previous best open-data result
- **100,000 training examples** and all ablation results published
- **Fully reproducible** \- anyone with enough GPUs can replicate the results

[arXiv →](https://arxiv.org/abs/2606.24855?ref=genaisecretsauce.com)

### Bigger AI Models Just Delay a Fundamental Problem - They Don't Solve It

**What this means for you:** The assumption that bigger AI models will keep getting better at learning new things turns out to be wrong - they just take longer to hit the wall.

- **Plasticity loss** (the inability to learn new information after mastering old information) follows a sublinear scaling law
- **Scaling delays the problem but never prevents it** \- even under ideal training conditions
- **Tested across models from 5M to 314M parameters** with consistent results

[arXiv →](https://arxiv.org/abs/2606.24752?ref=genaisecretsauce.com)

### Self-Recognition Training Can Prevent and Fix AI Identity Confusion

**What this means for you:** A new technique stops AI models from developing harmful behaviors after fine-tuning by stabilizing their sense of identity.

- **Emergent misalignment** occurs when fine-tuning on benign data disrupts the model's identity representation
- **Self-recognition finetuning** both prevents and reverses this failure mode
- **Tested on GPT-4.1, Qwen2.5-32B, and Seed-OSS-36B** \- works across model families

[arXiv →](https://arxiv.org/abs/2606.23700?ref=genaisecretsauce.com)

Business & Industry

## Business & Industry

### OpenAI Enters the Chip Business

- **Jalapeño is OpenAI's first custom ASIC** \- purpose-built for inference, co-developed with Broadcom
- **Nine-month development cycle** from design to tape-out, believed to be the fastest ever for a chip of this complexity
- **Microsoft purchasing \~40% of initial production** \- the largest single customer commitment

[Techcrunch →](https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom?ref=genaisecretsauce.com)

### Databricks Targets $175 Billion Valuation

- **Open-sourced Omnigent** \- a meta-harness for enterprise AI agents
- **400+ pull requests within days of launch** \- the fastest community adoption of any Databricks open-source project
- **Positioning as the "operating system for enterprise agents"** \- a direct challenge to cloud providers

[Latent Space →](https://www.latent.space/p/databricks?ref=genaisecretsauce.com)

### AI Economists Are Here: Labs Hiring Philosophers in Droves

- **The Economist reports major AI labs are hiring philosophers** to work on alignment, safety, and value specification
- **The demand reflects a shift** from pure engineering to interdisciplinary teams
- **Philosophy departments are seeing a talent drain** to industry for the first time

Education

## GenAI in Education

### The Mythos/Fable Crisis Reshapes AI Governance Debates in Academia

**What this means for you:** The federal shutdown of Anthropic's most capable models is forcing universities to confront what happens when AI tools they depend on disappear overnight.

Bryan Alexander's analysis traces the political and institutional implications of the Commerce Department's June 12 directive. The revelation that Mythos breached "almost all" NSA classified systems in hours during a red-team exercise has transformed what was an export control debate into a broader question about who should control access to powerful AI.

[Aiandacademia →](https://aiandacademia.substack.com/p/on-the-epic-and-still-developing)

### AI-Generated Job Applications Are Erasing Candidates, Not Enhancing Them

Simon Willison amplifies Tom MacWright's observation that applicants now submit entirely AI-generated materials - resumes, portfolios, GitHub repos, even individual commit messages. The paradox: the attempt to appear more professional produces applications indistinguishable from every other AI-generated application.

[Simon Willison →](https://simonwillison.net/2026/Jun/24/tom-macwright/?ref=genaisecretsauce.com#atom-everything)

### LeetCode Persists Because It Tests Scalability, Not Prediction

NeetCode (creator of NeetCode.io) told The Pragmatic Engineer that coding interviews persist not because they predict job performance but because they test whether a candidate can think about systems at scale. Google has restarted onsite whiteboard interviews specifically to prevent AI-assisted cheating.

[Newsletter →](https://newsletter.pragmaticengineer.com/p/tech-interviews-with-neetcode?ref=genaisecretsauce.com)

Surprising

## Surprising & Under-the-Radar

### AI Models Drop Their Scientific Rigor the Moment You Ask for Advice

LLMs maintain proper causal reasoning 91-100% of the time in academic contexts. But when a user asks for practical advice, that number crashes to 0.5-18%. A one-line correction prompt ("ensure your recommendations are supported by causal evidence") restores it to 71-100%. Tested on Claude Sonnet 4.6, Claude Opus 4.7, GPT 5.5, and Gemini 3.1 Pro.

[arXiv →](https://arxiv.org/abs/2606.24370?ref=genaisecretsauce.com)

### LLM Mental Health Safeguards Fail Completely for Eating Disorders

A follow-up audit across six proprietary LLMs and 16 DSM-5 conditions found that suicide and self-harm safeguards hold reliably. But eating disorders, substance use disorder, and major depressive disorder showed failure rates up to 100% under adversarial prompting. The safety net has holes in exactly the conditions where vulnerable users are most likely to seek help.

[arXiv →](https://arxiv.org/abs/2606.23884?ref=genaisecretsauce.com)

### Physicists Say LLM Scaling Is Thermodynamically Unsustainable

Physicists from University College London applied thermodynamic and fluid-turbulence reasoning to LLM scaling laws and concluded the scaling exponents are "too small to be sustainable" from an energy perspective. Their argument: the diminishing returns are not just an engineering challenge but a physical constraint.

[arXiv →](https://arxiv.org/abs/2606.24504?ref=genaisecretsauce.com)

### LLM Agent Societies Spontaneously Develop Social Hierarchies

Give LLM agents emotional states, identities, and social memory, then let them interact at scale. Five complex social phenomena emerge without being programmed: authority stratification, coalition formation, emotional contagion, norm enforcement, and reputation systems. Published in Findings of ACL 2026.

[arXiv →](https://arxiv.org/abs/2606.23764?ref=genaisecretsauce.com)

### Fender Is Suing Over Guitar Shapes - and Losing in Europe

Thomann, Europe's largest music retailer, filed legal action against Fender's attempt to monopolize the Stratocaster body shape through cease-and-desist campaigns. Not AI-related, but the intellectual property dynamics mirror debates about AI-generated content and design ownership.

[Thomann →](https://www.thomann.de/blog/en/inside/thomann-takes-legal-action-against-fender?ref=genaisecretsauce.com)

Worth Watching

## Signals to Track

[![Worth Watching](https://genaisecretsauce.com/content/images/2026/06/section-worth-watching-2026-06-24-1.png)](https://genaisecretsauce.com/content/images/2026/06/section-worth-watching-2026-06-24-1.png) 

01

### Cryptographic Proof Could Make AI Agents Auditable by Default

Every action an AI agent takes could come with a mathematical proof that it followed the rules.

A new paper proposes attaching independently verifiable cryptographic certificates to every agent action, proving compliance with formally specified policies. The approach translates policy requirements into logical predicates and generates proofs using zero-knowledge systems. If this scales, it could resolve the "trust but verify" problem for autonomous AI systems.

[arXiv →](https://arxiv.org/abs/2606.23768?ref=genaisecretsauce.com)

02

### FP8 Tensor Core Tricks Could Unlock the Next Generation of GPUs

NVIDIA's newest chip has a hidden limitation - and researchers just found a software workaround.

The B300 Graphics Processing Unit (GPU) has 30x less native FP64 throughput than the B200, which would cripple scientific computing. A new paper routes calculations through FP8 tensor cores using mathematical reformulations, achieving FP64-equivalent precision at FP8 speed. This matters because it determines whether the newest hardware can serve both AI and scientific workloads.

[arXiv →](https://arxiv.org/abs/2606.23698?ref=genaisecretsauce.com)

03

### Bayesian Control Could Make Coding Agents Dramatically Cheaper

Instead of running every test on every change, let probability decide when to stop.

A paper reframes coding agent orchestration as Bayesian hypothesis testing: maintain a probabilistic belief about whether code is correct, then dynamically decide whether to gather more evidence or ship. The cost-performance tradeoff outperforms fixed deterministic pipelines.

[arXiv →](https://arxiv.org/abs/2606.24453?ref=genaisecretsauce.com)

04

### Far-Field Speech Recognition Has Its First Real Benchmark

AI speech recognition works great - until you're more than a few feet from the microphone.

The new FFASR Leaderboard reveals that error rates for far-field conditions (reverberant rooms, background noise, distance) are "several times higher" than near-field benchmarks. This is the gap between demo and deployment for every smart speaker, conference room, and voice-controlled device.

[Hugging Face: blog/ffasr-leaderboard →](https://huggingface.co/blog/ffasr-leaderboard?ref=genaisecretsauce.com)

05

### Multi-Agent Memory Sharing Has Four Fundamental Failure Modes

When AI agents share a knowledge base, four things go wrong: leakage, staleness, contradiction, and lost provenance.

A formalization of the "fleet-memory problem" identifies these failure modes and proposes system-level primitives (access control, versioning, conflict resolution, attribution) to address them. As multi-agent systems move from research to production, this taxonomy will define the engineering requirements.

[arXiv →](https://arxiv.org/abs/2606.24535?ref=genaisecretsauce.com)

GitHub Trending

## Top Repos Today

#1

### [calesthio/OpenMontage](https://github.com/calesthio/OpenMontage?ref=genaisecretsauce.com)

Rank yesterday: #1 - Holding steady ➡

⭐ **Stars today:** +3,703 · 📦 **Total:** 19,211  
📜 **License:** AGPL-3.0 · 👤 **By:** Individual developer  
🎯 **Time to value:** 15 minutes

**What it is:** An open-source agentic video production platform that orchestrates 12 production workflows, 52 tools, and 400+ agent skills for end-to-end video creation. Previously [covered June 20](https://genaisecretsauce.com/genai-secret-sauce-daily-digest-2026-06-20/) and [June 23](https://genaisecretsauce.com/genai-secret-sauce-daily-digest-2026-06-23/). **Why you'd want it:** Full video production pipeline from research to final composition in a single system, without stitching separate tools.

| ✓ Pros                                        | ✗ Cons                             |
| --------------------------------------------- | ---------------------------------- |
| End-to-end pipeline with 400+ skills          | AGPL license limits commercial use |
| Integrates multiple AI image/video generators | Requires significant GPU resources |
| Active community (19K+ stars in days)         | Solo maintainer - bus factor of 1  |

[GitHub - calesthio/OpenMontage: World’s first open-source, agentic video production system. 12 pipelines, 52 tools, 500+ agent skills. Turn your AI coding assistant into a full video production studio.World’s first open-source, agentic video production system. 12 pipelines, 52 tools, 500+ agent skills. Turn your AI coding assistant into a full video production studio. - calesthio/OpenMontage![](https://genaisecretsauce.com/content/images/icon/pinned-octocat-093da3e6fa40-b0fc6990-bc7a-4ada-bc0d-d631ce25dc84.svg)calesthioGitHub![](https://genaisecretsauce.com/content/images/thumbnail/645f0bc1-450d-4791-99b8-6102b4bb9f3d-f84e2fe0-7dfe-486d-b223-64568c4bcea0)](https://github.com/calesthio/OpenMontage?ref=genaisecretsauce.com)

#2

### [ZhuLinsen/daily\_stock\_analysis](https://github.com/ZhuLinsen/daily%5Fstock%5Fanalysis?ref=genaisecretsauce.com)

Rank yesterday: #2 - Holding steady ➡

⭐ **Stars today:** +1,461 · 📦 **Total:** 48,425  
📜 **License:** MIT · 👤 **By:** Individual developer  
🎯 **Time to value:** 10 minutes

**What it is:** An LLM-powered stock analysis system that automatically analyzes Chinese, Hong Kong, U.S., Japanese, and Korean equities daily. It fetches real-time data and news, generates AI-driven reports, and delivers them via WeChat, Telegram, or email. **Why you'd want it:** A self-running daily stock briefing across five markets with AI-generated analysis each morning.

| ✓ Pros                            | ✗ Cons                                     |
| --------------------------------- | ------------------------------------------ |
| Multi-market coverage (5 markets) | Analysis quality limited by underlying LLM |
| Automated daily delivery          | Requires API keys for market data          |
| MIT license for commercial use    | Not a substitute for professional advice   |

[GitHub - ZhuLinsen/daily\_stock\_analysis: LLM 驱动的多市场股票智能分析系统：多源行情、实时新闻、决策看板与自动推送，支持零成本定时运行。 LLM-powered multi-market stock analysis system with multi-source market data, real-time news, decision dashboard, automated notifications, and cost-free scheduled runs.LLM 驱动的多市场股票智能分析系统：多源行情、实时新闻、决策看板与自动推送，支持零成本定时运行。 LLM-powered multi-market stock analysis system with multi-source market data, real-time news, decision dashboard, automated notifications, and cos…![](https://genaisecretsauce.com/content/images/icon/pinned-octocat-093da3e6fa40-5b6d0f45-6c94-4ba3-87aa-7513beee75ec.svg)ZhuLinsenGitHub![](https://genaisecretsauce.com/content/images/thumbnail/2a028e99-3b90-4290-8381-d86fffbf3332-72cf2a10-3c50-480e-bf3c-f70ff23e89ad)](https://github.com/ZhuLinsen/daily%5Fstock%5Fanalysis?ref=genaisecretsauce.com)

#3

### [NousResearch/hermes-agent](https://github.com/NousResearch/hermes-agent?ref=genaisecretsauce.com)

Rank yesterday: #3 - Holding steady ➡

⭐ **Stars today:** +1,174 · 📦 **Total:** 202,010  
📜 **License:** MIT · 👤 **By:** Research lab (Nous Research)  
🎯 **Time to value:** 5 minutes

**What it is:** A self-improving AI assistant with an integrated learning loop - it autonomously creates and refines its own skills from experience. Runs across terminal, Telegram, Discord, and Slack with persistent memory across sessions. **Why you'd want it:** An agent that genuinely learns from use rather than resetting each session, with multi-platform reach.

| ✓ Pros                          | ✗ Cons                                      |
| ------------------------------- | ------------------------------------------- |
| Self-improving skill creation   | Learning loop quality varies by task domain |
| Persistent cross-session memory | High star count suggests community traction |
| Runs on consumer hardware       | Privacy implications of persistent memory   |

[GitHub - NousResearch/hermes-agent: The agent that grows with youThe agent that grows with you. Contribute to NousResearch/hermes-agent development by creating an account on GitHub.![](https://genaisecretsauce.com/content/images/icon/pinned-octocat-093da3e6fa40-f6a82828-2569-4839-b219-d3e1704f148f.svg)NousResearchGitHub![](https://genaisecretsauce.com/content/images/thumbnail/hermes-agent-92546b98-c908-4269-b770-1f325ea4c155)](https://github.com/NousResearch/hermes-agent?ref=genaisecretsauce.com)

#4

### [interviewstreet/hiring-agent](https://github.com/interviewstreet/hiring-agent?ref=genaisecretsauce.com)

Rank yesterday: New entry 🆕

⭐ **Stars today:** +152 · 📦 **Total:** 2,114  
📜 **License:** MIT · 👤 **By:** Company (HackerRank)  
🎯 **Time to value:** 10 minutes

**What it is:** HackerRank's open-source resume screening agent. Processes PDFs through extraction, GitHub enrichment, and structured scoring with fairness constraints and evidence citations. **Why you'd want it:** Consistent, reproducible first-pass resume screening with bias-reduction guardrails, from a company that processes millions of technical assessments.

| ✓ Pros                             | ✗ Cons                                        |
| ---------------------------------- | --------------------------------------------- |
| Fairness-constrained scoring       | AI resume screening carries legal risks       |
| GitHub contribution enrichment     | Favors candidates with public GitHub profiles |
| HackerRank backing and maintenance | Still requires human review of top candidates |

[GitHub - interviewstreet/hiring-agent: AI agent to evaluate and score resumes.AI agent to evaluate and score resumes. Contribute to interviewstreet/hiring-agent development by creating an account on GitHub.![](https://genaisecretsauce.com/content/images/icon/pinned-octocat-093da3e6fa40-34b7c05e-ecf8-4bd8-a610-c47401e77592.svg)interviewstreetGitHub![](https://genaisecretsauce.com/content/images/thumbnail/hiring-agent-30e9f40f-1d43-4921-bf28-a2eccb105b06)](https://github.com/interviewstreet/hiring-agent?ref=genaisecretsauce.com)

#5

### [JCodesMore/ai-website-cloner-template](https://github.com/JCodesMore/ai-website-cloner-template?ref=genaisecretsauce.com)

Rank yesterday: New entry 🆕

⭐ **Stars today:** +693 · 📦 **Total:** 19,260  
📜 **License:** MIT · 👤 **By:** Individual developer  
🎯 **Time to value:** 5 minutes

**What it is:** A template that uses AI coding agents to reverse-engineer any website into a modern Next.js codebase. Extracts design tokens, assets, and component specifications, then reconstructs each section with parallel AI builders. **Why you'd want it:** Cuts days of manual design-to-code work to minutes for prototyping, migration, or competitive analysis.

| ✓ Pros                                  | ✗ Cons                                           |
| --------------------------------------- | ------------------------------------------------ |
| Works with Claude Code, Copilot, others | Ethical and legal grey area for cloning          |
| Parallel builder architecture           | Output quality depends on source site complexity |
| MIT license                             | May miss interactive/dynamic elements            |

[GitHub - JCodesMore/ai-website-cloner-template: Clone any website with one command using AI coding agentsClone any website with one command using AI coding agents - JCodesMore/ai-website-cloner-template![](https://genaisecretsauce.com/content/images/icon/pinned-octocat-093da3e6fa40-565e6152-5172-4c57-98c7-aa6c758404ff.svg)JCodesMoreGitHub![](https://genaisecretsauce.com/content/images/thumbnail/ai-website-cloner-template-0a22d53b-439a-46a1-810e-ba7b4016ea68)](https://github.com/JCodesMore/ai-website-cloner-template?ref=genaisecretsauce.com)

#6

### [google-labs-code/design.md](https://github.com/google-labs-code/design.md?ref=genaisecretsauce.com)

Rank yesterday: Holding steady ➡

⭐ **Stars today:** +504 · 📦 **Total:** 17,268  
📜 **License:** Apache-2.0 · 👤 **By:** Company (Google Labs)  
🎯 **Time to value:** 10 minutes

**What it is:** A format specification that combines machine-readable design tokens with human-readable prose, so AI coding agents can consistently apply a project's visual identity. Includes a validator, change-detection tool, and exporters for Tailwind CSS. **Why you'd want it:** Stops AI coding agents from guessing colors, spacing, and typography - gives them an authoritative design system source of truth.

| ✓ Pros                                | ✗ Cons                                         |
| ------------------------------------- | ---------------------------------------------- |
| Google Labs backing                   | Requires upfront design token authoring        |
| Tailwind CSS and W3C format exporters | Only useful if your team uses AI coding agents |
| Change detection across versions      | Early-stage, format may evolve                 |

[GitHub - google-labs-code/design.md: A format specification for describing a visual identity to coding agents. DESIGN.md gives agents a persistent, structured understanding of a design system.A format specification for describing a visual identity to coding agents. DESIGN.md gives agents a persistent, structured understanding of a design system. - google-labs-code/design.md![](https://genaisecretsauce.com/content/images/icon/pinned-octocat-093da3e6fa40-a3a16757-4c94-4591-ade1-536236db7dd3.svg)google-labs-codeGitHub![](https://genaisecretsauce.com/content/images/thumbnail/design-feb7edd2-9497-4d52-8312-47f654819f47.md)](https://github.com/google-labs-code/design.md?ref=genaisecretsauce.com)

#7

### [stablyai/orca](https://github.com/stablyai/orca?ref=genaisecretsauce.com)

Rank yesterday: Holding steady ➡

⭐ **Stars today:** +387 · 📦 **Total:** 6,736  
📜 **License:** MIT · 👤 **By:** Company (Stably AI)  
🎯 **Time to value:** 5 minutes

**What it is:** An Agent Development Environment (ADE) that runs multiple AI coding agents simultaneously in isolated git worktrees so they don't conflict. Supports Claude Code, Codex, and others with desktop and mobile interfaces. **Why you'd want it:** Parallelizes AI coding by running 10+ agents at once without merge conflicts.

| ✓ Pros                                    | ✗ Cons                                       |
| ----------------------------------------- | -------------------------------------------- |
| Isolated worktrees prevent conflicts      | Resource-intensive with many agents          |
| Multi-agent support (Claude, Codex, etc.) | Coordination between agents is manual        |
| GitHub and Linear integrations            | Requires understanding of git worktree model |

[GitHub - stablyai/orca: Orca is the ADE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop and mobile.Orca is the ADE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop and mobile. - stablyai/orca![](https://genaisecretsauce.com/content/images/icon/pinned-octocat-093da3e6fa40-3e71586c-b199-47a8-987b-bd819ee3e102.svg)stablyaiGitHub![](https://genaisecretsauce.com/content/images/thumbnail/orca-e0393ab4-7c7b-4025-91ab-72aee157d680)](https://github.com/stablyai/orca?ref=genaisecretsauce.com)

#8

### [revfactory/harness](https://github.com/revfactory/harness?ref=genaisecretsauce.com)

Rank yesterday: Holding steady ➡

⭐ **Stars today:** +274 · 📦 **Total:** 7,712  
📜 **License:** Apache-2.0 · 👤 **By:** Individual developer  
🎯 **Time to value:** 10 minutes

**What it is:** A meta-skill for Claude Code that automatically designs multi-agent teams from a plain-language domain description. Selects from six architectural patterns (Pipeline, Fan-out/Fan-in, Expert Pool, etc.) and generates coordinated agent systems with custom skills. **Why you'd want it:** Describes your problem domain in plain English and gets a complete multi-agent architecture in return.

| ✓ Pros                       | ✗ Cons                                      |
| ---------------------------- | ------------------------------------------- |
| Six architectural patterns   | Claude Code-specific                        |
| Plain language to agent team | Generated architectures may need refinement |
| Apache-2.0 license           | Relatively niche use case                   |

[GitHub - revfactory/harness: A meta-skill that designs domain-specific agent teams, defines specialized agents, and generates the skills they use.A meta-skill that designs domain-specific agent teams, defines specialized agents, and generates the skills they use. - revfactory/harness![](https://genaisecretsauce.com/content/images/icon/pinned-octocat-093da3e6fa40-290922c5-9d57-4c4c-9c42-acec4f76f2aa.svg)revfactoryGitHub![](https://genaisecretsauce.com/content/images/thumbnail/harness-550c1363-de11-46d4-9606-810089a34930)](https://github.com/revfactory/harness?ref=genaisecretsauce.com)

HuggingFace Trending

## Top Models Today

#1

### [deepseek-ai/DeepSeek-V4-Pro](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro?ref=genaisecretsauce.com)

Previously [covered June 21](https://genaisecretsauce.com/genai-secret-sauce-daily-digest-2026-06-21/) \- DeepSeek V4-Pro launched with 1.6 trillion parameters on Huawei chips.

📥 **Downloads (30d):** 2.05M · 📜 **License:** DeepSeek  
👤 **By:** DeepSeek AI · 🎯 **Task:** text-generation  
📐 **Size:** 862B (49B active)

**What it is:** A Mixture-of-Experts language model with 1.6 trillion total parameters but only 49 billion active per query. Features hybrid attention and a 1-million-token context window. **Why you'd want it:** The most capable open-weight model currently available, with Mixture of Experts (MoE) efficiency making it practical to deploy despite its massive size.

| ✓ Pros                     | ✗ Cons                                  |
| -------------------------- | --------------------------------------- |
| 1M token context window    | Requires significant infrastructure     |
| Only 49B active parameters | DeepSeek license may restrict some uses |
| 2M+ downloads in 30 days   | Chinese origin may face export controls |

[deepseek-ai/DeepSeek-V4-Pro · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.![](https://genaisecretsauce.com/content/images/icon/favicon-67eb90a3-8c98-47eb-b313-041e4994bed4.ico)![](https://genaisecretsauce.com/content/images/thumbnail/DeepSeek-V4-Pro-bd68370c-058a-46d3-8061-e8c77361e676.png)](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro?ref=genaisecretsauce.com)

#2

### [zai-org/GLM-5.2](https://huggingface.co/zai-org/GLM-5.2?ref=genaisecretsauce.com)

Previously [covered June 21](https://genaisecretsauce.com/genai-secret-sauce-daily-digest-2026-06-21/) \- GLM-5.2 beat GPT-5.5 on multi-hour coding benchmarks.

📥 **Downloads (30d):** 57.2K · 📜 **License:** Open  
👤 **By:** Z.ai · 🎯 **Task:** text-generation  
📐 **Size:** 753B

**What it is:** A fully open-source 753B language model with a 1M-token context window, advanced coding capabilities, and adjustable reasoning depth. **Why you'd want it:** The strongest fully open-source model available, with no usage restrictions.

| ✓ Pros                                 | ✗ Cons                             |
| -------------------------------------- | ---------------------------------- |
| Fully open source with no restrictions | 753B requires multi-GPU deployment |
| Adjustable reasoning depth             | Newer model, smaller community     |
| Beat GPT-5.5 on coding benchmarks      | Download count still low           |

[zai-org/GLM-5.2 · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.![](https://genaisecretsauce.com/content/images/icon/favicon-12c91604-875b-4e5f-bd3c-f3c351e5015b.ico)![](https://genaisecretsauce.com/content/images/thumbnail/GLM-5.2-ded2fad7-8139-4ff3-a9a3-7a3302488eac.png)](https://huggingface.co/zai-org/GLM-5.2?ref=genaisecretsauce.com)

#3

### [MiniMaxAI/MiniMax-M3](https://huggingface.co/MiniMaxAI/MiniMax-M3?ref=genaisecretsauce.com)

Previously [covered June 20](https://genaisecretsauce.com/genai-secret-sauce-daily-digest-2026-06-20/) \- MiniMax-M3 launched as a 427B open multimodal model.

📥 **Downloads (30d):** 143K · 📜 **License:** MiniMax Open  
👤 **By:** MiniMax AI · 🎯 **Task:** multimodal  
📐 **Size:** 428B (23B active)

**What it is:** A native multimodal MoE model with 428B total and 23B active parameters, processing text, images, audio, and video with a 1M-token context. **Why you'd want it:** The most capable open multimodal model, processing four modalities in a single model.

| ✓ Pros                       | ✗ Cons                                   |
| ---------------------------- | ---------------------------------------- |
| Four modalities in one model | 23B active is still GPU-intensive        |
| 1M token context             | MiniMax license may restrict some uses   |
| 143K downloads show traction | Chinese origin and licensing uncertainty |

[MiniMaxAI/MiniMax-M3 · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.![](https://genaisecretsauce.com/content/images/icon/favicon-aef9f925-c0d4-48c7-991a-b53c47a9e8db.ico)![](https://genaisecretsauce.com/content/images/thumbnail/MiniMax-M3-2b6a4309-6134-4e02-9195-60fb57829d2c.png)](https://huggingface.co/MiniMaxAI/MiniMax-M3?ref=genaisecretsauce.com)

#4

### [google/diffusiongemma-26B-A4B-it](https://huggingface.co/google/diffusiongemma-26B-A4B-it?ref=genaisecretsauce.com)

New entry - first appearance on trending.

📥 **Downloads (30d):** 1.04M · 📜 **License:** Gemma  
👤 **By:** Google DeepMind · 🎯 **Task:** multimodal generation  
📐 **Size:** 26B (3.8B active)

**What it is:** Google DeepMind's multimodal discrete-diffusion model that generates text and images together. Uses a Mixture-of-Experts design with only 3.8 billion active parameters. **Why you'd want it:** Generates multimodal content with a fraction of the compute of comparable models, from a major lab with a permissive license.

| ✓ Pros                      | ✗ Cons                                    |
| --------------------------- | ----------------------------------------- |
| Only 3.8B active parameters | Gemma license terms                       |
| 1M+ downloads in 30 days    | Discrete diffusion is newer, less tooling |
| Google DeepMind backing     | May lag behind specialized image models   |

[google/diffusiongemma-26B-A4B-it · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.![](https://genaisecretsauce.com/content/images/icon/favicon-15caf65f-9ec2-406d-b5b1-504971043f96.ico)![](https://genaisecretsauce.com/content/images/thumbnail/diffusiongemma-26B-A4B-it-6f004b61-8558-46f0-8a1e-03d17f8c0d81.png)](https://huggingface.co/google/diffusiongemma-26B-A4B-it?ref=genaisecretsauce.com)

#5

### [baidu/Unlimited-OCR (Optical Character Recognition)](https://huggingface.co/baidu/Unlimited-OCR?ref=genaisecretsauce.com)

New entry - first appearance on trending.

📥 **Downloads (30d):** 45.7K · 📜 **License:** Apache 2.0  
👤 **By:** Baidu · 🎯 **Task:** OCR/document processing  
📐 **Size:** 3B

**What it is:** Baidu's 3B vision-language model that parses entire multi-page PDFs in a single pass with no length limits, extracting text, tables, and layout information. **Why you'd want it:** Process any PDF regardless of length in one shot - useful for legal documents, research papers, and financial reports.

| ✓ Pros                 | ✗ Cons                                 |
| ---------------------- | -------------------------------------- |
| No page limit on PDFs  | Accuracy on complex layouts unverified |
| Apache 2.0 license     | 3B model may struggle with handwriting |
| Single-pass processing | Limited community documentation        |

[baidu/Unlimited-OCR · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.![](https://genaisecretsauce.com/content/images/icon/favicon-096cb376-4a4a-4e3a-9f02-b90b2ad900dd.ico)![](https://genaisecretsauce.com/content/images/thumbnail/Unlimited-OCR-d69566e7-807b-408f-90cc-1c967375ac7d.png)](https://huggingface.co/baidu/Unlimited-OCR?ref=genaisecretsauce.com)

#6

### [nvidia/LocateAnything-3B](https://huggingface.co/nvidia/LocateAnything-3B?ref=genaisecretsauce.com)

Continuing to trend (previously covered).

📥 **Downloads (30d):** 359K · 📜 **License:** NVIDIA  
👤 **By:** NVIDIA · 🎯 **Task:** visual grounding  
📐 **Size:** 3B

**What it is:** NVIDIA's model for locating any object or UI element in an image from a text description. Outputs bounding boxes for arbitrary visual targets. **Why you'd want it:** Point at anything in any image using words - useful for automated testing, accessibility, and UI automation.

| ✓ Pros                              | ✗ Cons                                |
| ----------------------------------- | ------------------------------------- |
| Text-to-bounding-box in any image   | NVIDIA license may restrict some uses |
| 359K downloads show strong adoption | 3B model needs GPU                    |
| Versatile across domains            | Accuracy varies with image complexity |

[nvidia/LocateAnything-3B · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.![](https://genaisecretsauce.com/content/images/icon/favicon-0abea9f7-8c4f-414a-adfa-54870dc963bb.ico)![](https://genaisecretsauce.com/content/images/thumbnail/LocateAnything-3B-28d3b3e5-1009-49c6-b9e0-c4ee03226f88.png)](https://huggingface.co/nvidia/LocateAnything-3B?ref=genaisecretsauce.com)

#7

### [microsoft/FastContext-1.0-4B-SFT](https://huggingface.co/microsoft/FastContext-1.0-4B-SFT?ref=genaisecretsauce.com)

Continuing to trend (previously covered).

📥 **Downloads (30d):** 4.81K · 📜 **License:** MIT  
👤 **By:** Microsoft · 🎯 **Task:** coding subagent  
📐 **Size:** 4B

**What it is:** Microsoft's 4B model designed specifically as a coding subagent that handles repository exploration for main coding agents, cutting context-gathering costs by 60%. **Why you'd want it:** A specialized model that makes your main coding agent cheaper by offloading the expensive exploration step.

| ✓ Pros                             | ✗ Cons                                      |
| ---------------------------------- | ------------------------------------------- |
| 60% cost reduction for exploration | Only useful paired with a main coding agent |
| MIT license                        | Small model may miss complex patterns       |
| Microsoft backing                  | Low download count suggests early adoption  |

[microsoft/FastContext-1.0-4B-SFT · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.![](https://genaisecretsauce.com/content/images/icon/favicon-5472c238-06ce-4d01-81ef-560334449197.ico)![](https://genaisecretsauce.com/content/images/thumbnail/FastContext-1.0-4B-SFT-6119214e-764a-4ce0-8c46-fe1d69c869fe.png)](https://huggingface.co/microsoft/FastContext-1.0-4B-SFT?ref=genaisecretsauce.com)

Product Hunt

## AI Launches Today

### [Propane](https://www.producthunt.com/products/propane?ref=genaisecretsauce.com)

Automatic customer context for product teams and agents

🔥 **Upvotes:** 437 · 👤 **By:** Propane team  
💰 **Pricing:** Not specified · 🏷 **Category:** Customer Intelligence

Propane automatically collects and unifies customer context from all your tools into one always-current view, accessible by both human product teams and AI agents. It solves the "data scattered across 12 tools" problem by maintaining a single customer truth layer. **Verdict:** Strong upvote count suggests real demand for unified customer context, especially as AI agents need structured customer data to be useful.

[Propane : Automatic customer context for product teams and agents | Product HuntPropane gives your product team and agents one connected, always-current view of your customers. Automatically collected from all your tools. Collaborate on a shared canvas. Commit straight to any coding or design agent. Secure, maintained, always on. You just build products people love.![](https://genaisecretsauce.com/content/images/icon/ph-favicon-brand-500-efe6fa5e-53ad-490c-886e-33318c324bf5.svg)John Acosta and Nicolai Harms and Anders T.Product Hunt![](https://genaisecretsauce.com/content/images/thumbnail/2259907b-26bf-46c8-8741-6ff6225c79f6-bfbaeace-7949-4936-9d53-629312131c8b.png)](https://www.producthunt.com/products/propane?ref=genaisecretsauce.com)

### [Tencent EdgeOne Makers](https://www.producthunt.com/products/tencent-edgeone-makers?ref=genaisecretsauce.com)

Ship AI agents like web apps, in minutes

🔥 **Upvotes:** 357 · 👤 **By:** Tencent  
💰 **Pricing:** Freemium · 🏷 **Category:** Developer Infrastructure

Edge network platform integrating CDN, DNS, WAF, and DDoS protection with AI agent hosting. Positions itself as Vercel for AI agents - deploy and scale agents at the edge. **Verdict:** Tencent's infrastructure backing gives this credibility, but the "ship agents like web apps" promise needs real-world testing.

[View on Product Hunt →](https://www.producthunt.com/products/tencent-edgeone-makers?ref=genaisecretsauce.com) 

### [Buy by Agentcard](https://www.producthunt.com/products/buy-by-agentcard?ref=genaisecretsauce.com)

Order DoorDash from Claude

🔥 **Upvotes:** 154 · 👤 **By:** Agentcard  
💰 **Pricing:** Free · 🏷 **Category:** AI Commerce

A debit card that enables AI agents to safely make online purchases. The first concrete product solving the "how do agents pay for things" problem. **Verdict:** The most provocative launch today. Agent-controlled spending is a real frontier, and a dedicated payment rail is probably the right approach.

[View on Product Hunt →](https://www.producthunt.com/products/buy-by-agentcard?ref=genaisecretsauce.com) 

### [FUTO Swipe](https://www.producthunt.com/products/futo-swipe?ref=genaisecretsauce.com)

Open models for on-device swipe typing

🔥 **Upvotes:** 125 · 👤 **By:** FUTO  
💰 **Pricing:** Free · 🏷 **Category:** On-Device AI

Small, open-source models purpose-built for accurate swipe keyboard typing that run entirely on-device. No cloud, no data collection. **Verdict:** A refreshing privacy-first approach in a category dominated by cloud-based keyboards that harvest typing data.

[FUTO Swipe: Open models for on-device swipe typing | Product HuntFUTO Swipe is a family of small, open models for accurate swipe typing. It includes a layout-agnostic encoder, a layout-specific decoder, and a lightweight context language model. The full system runs efficiently on-device with a very small footprint, and FUTO has also released the 1 million swipe dataset used to train it.![](https://genaisecretsauce.com/content/images/icon/ph-favicon-brand-500-3adcb85b-26e0-4498-8205-8592ec52da08.svg)Product Hunt![](https://genaisecretsauce.com/content/images/thumbnail/86323265-698f-4e5c-b4e1-f0c3fcdadbd0-af3deffa-9283-480d-9801-01867972898b.jpeg)](https://www.producthunt.com/products/futo-swipe?ref=genaisecretsauce.com)

API Pricing

## Snapshot

Provider

Model

Input $/1M

Output $/1M

Context

Anthropic

Claude Fable 5

$10.00

$50.00

1M

Anthropic

Claude Opus 4.8

$5.00

$25.00

1M

Anthropic

Claude Sonnet 4.6

$3.00

$15.00

1M

Anthropic

Claude Haiku 4.5

$1.00

$5.00

200K

OpenAI

GPT-5.5

$5.00

$30.00

Long

OpenAI

GPT-5.5 Pro

$30.00

$180.00

Long

OpenAI

GPT-5.4 Mini

$0.75

$4.50

Short

OpenAI

GPT-5.4 Nano

$0.20

$1.25

Short

Google

Gemini 3.5 Flash

$1.50

$9.00

N/A

Google

Gemini 2.5 Flash

$0.30

$2.50

1M

Google

Gemini 2.5 Flash-Lite

$0.10

$0.40

N/A

Groq

Llama 4 Scout

$0.11

$0.34

128K

Groq

Llama 3.1 8B Instant

$0.05

$0.08

128K

**What this means:** The price floor continues to fall. Groq's Llama 3.1 8B at $0.05/$0.08 per million tokens is 200x cheaper than OpenAI's GPT-5.5 Pro. Google's Flash-Lite at $0.10/$0.40 is the cheapest offering from a major frontier lab. The gap between premium and commodity tiers is now over 100x - the widest it has ever been.  
  
arXiv Paper of the Day

## CompressKV: Semantic-Retrieval-Guided KV-Cache Compression for Resource-Efficient Long-Context LLM Inference

Xiaolin Lin, Jingcun Wang, Olga Kondrateva, Yiyu Shi, Bing Li, Grace Li Zhang · arXiv:2606.24467

**What it claims:** By identifying "Semantic Retrieval Heads" - the specific attention heads responsible for locating contextually critical tokens - CompressKV concentrates the entire key-value cache budget on those heads and discards the rest. Combined with per-layer eviction-error budgeting, it retains nearly all model accuracy while slashing memory by 30-97x.  
  
**Key finding:** **97% of full-cache performance at 3% memory cost on LongBench QA. 90% accuracy with just 0.7% KV storage on Needle-in-a-Haystack.**  
  
**Why practitioners should care:** KV-cache memory is the binding constraint for long-context LLM batching in production today. A training-free, drop-in compression layer that cuts memory by 30-97x with negligible accuracy loss means larger batch sizes, longer supported contexts, or dramatically lower GPU costs on existing hardware. No model changes required.  
  
[Read on arXiv →](https://arxiv.org/abs/2606.24467?ref=genaisecretsauce.com)

GenAI Secret Sauce Daily Digest · 2026-06-24

×

Click anywhere or press ESC to close