Search the archive

EMPATHY AND DELIGHT MEAN NOTHING WHEN THE SOFTWARE IS DISRESPECTFUL

Empathy and delight haven't improved software because respect is missing. Products remove user control and push work onto users through rigid flows and forms. This starts with designers who overestimate their role and scales through systems that prioritize efficiency over agency. LLMs extend this by defaulting to doing tasks for users, generating low-value output that increases review burden and rewards volume over judgment. The result is bloated features and worse experiences. Respectful software gives users control, keeps useful friction, and communicates clearly.

MILLIONS OF AMERICANS TAKE ON DEBT AND DRAIN SAVINGS TO AFFORD FOOD

About 20% of working-age adults used long-term savings to pay for groceries in the past year.

EU COMMISSION PRELIMINARILY FINDS META APPS ADDICTIVE

Meta may face fines and forced product changes in Europe after the EU ruled that Facebook and Instagram's addictive features likely violate its laws.

THIS K-BEAUTY BRAND IS COLLAPSING THE MARKETING FUNNEL TO GROW ITS US CUSTOMER BASE

Sulwhasoo is collapsing the funnel by pairing CTV and DOOH with retargeting.

THE 4 S'S OF YOUTUBE SUCCESS

YouTube growth is attributed to subscribers, suggested, search, and shorts.

THE REVERSE INFORMATION PARADOX

Companies risk surrendering proprietary knowledge as their employees use and improve external AI systems. To retain that value, enterprises must control their data, feedback, evaluations, memory, and learning infrastructure while remaining independent from any single model provider.

THE MOST HUMAN TECHNOLOGY EVER MADE

AI's greatest promise is not simply saving time. It gives more people the ability to create, experiment, and turn their ideas into reality, making individuality more valuable than technical expertise, capital, or permission.

TASTE CANNOT BE DELEGATED

AI can generate endless design options, but it cannot determine which one best serves a product's ambition and context. Strong design still requires a responsible leader who can apply taste, resolve tradeoffs, and confidently choose a direction.

FIVE YEARS BUILDING THE CONTEXT LAYER - AND WHY IT BELONGS IN THE AGENTIC LOOP

Graphlit built a powerful production platform but failed to become a scalable business because it lacked repeatable distribution and a clear, outcome-driven market position. Its core insight remains: agents need a portable context layer that stays separate from the runtime while supporting every stage of the work with durable organizational knowledge.

HOW TO COLLECT PRODUCT FEEDBACK WHEN YOUR AI GIVES EVERY USER A DIFFERENT ANSWER

AI outputs vary from one interaction to another, so traditional feedback methods often provide an incomplete picture. Product teams should focus on users' next actions, review a representative range of real outputs, and capture the context behind poor results.

AI'S BIGGEST WINNERS HAVE THE LOWEST MARGINS

Low-margin businesses could become some of AI's biggest winners because modest cost reductions can dramatically increase profits. The greatest opportunity lies in embedding AI into existing operations to remove coordination work without requiring employees to change how they work.

INCUBATE, COMPOUND, REFINANCE, LIQUIDATE

AI's value depends on whether it moves software assets toward healthier economics, not simply whether it makes teams faster.

THE FORGOTTEN MEASURE OF DATA QUALITY: DECISION QUALITY

Clean, reliable data has little economic value unless it leads to better decisions. Data teams should define the decisions each product is meant to improve and measure that impact alongside traditional quality metrics.

WITH AI AGENTS, ATTENTION IS THE NEW BOTTLENECK

AI makes execution cheap, so the advantage shifts to teams that manage attention and judgment well.

META ADMITS ITS FIRST ‘SUPERINTELLIGENCE' WAS TOO STUPID TO SURVIVE FOR THREE DAYS

Meta pulled Muse Image after its lack of clear consent controls sparked immediate privacy and misuse concerns.

Try Bob for free

APPLE SUES OPENAI, ACCUSING IT OF STEALING COMPANY SECRETS

Apple has accused OpenAI of stealing secrets about products still in development. OpenAI's new hardware business allegedly asked job candidates from Apple to share details about secret projects and to bring device components and prototypes to their interviews. It used the information to approach at least one of Apple's manufacturing partners, asking them to demonstrate Apple's technique for finishing metal on its devices. Apple is seeking an injunction that would prevent OpenAI from possessing, using, or sharing its trade secrets, as well as an order requiring OpenAI to return Apple's intellectual property.

APPLE'S M6, M7, AND M8 CHIPS SHOW HOW AI IS RESHAPING THE COMPANY

AI is no longer just another feature that Apple's chips need to support. It is now shaping how Apple's products are designed and when they are shipped. The Apple car project was not a futile exercise. The AI hardware effort developed for the vehicle now powers Macs and AI servers. What was seen as one of Apple's costliest failures may actually have been one of its most consequential technology investments.

HOME ROBOTS ALREADY WALK. 1X'S NEW HANDS TRY TO SOLVE THE PART THAT ACTUALLY MATTERS

1X's new robotic hands have 25 backdrivable joints that give way when pushed rather than staying rigid. The skin's sensors read both pressure and sideways movement across the fingers, allowing it to notice a glass starting to slip. Factory robots use grippers because they work with parts placed exactly at the same spot every time, but homes have a much larger variation in object type and placement. 1X's robotic hands can bend past a human range and wrap around awkward shapes and are ready for chores.

CHINA RECOVERED ITS FIRST REUSABLE ROCKET AND SHOWED A NEW WAY TO DO IT

China has demonstrated its first-ever controlled rocket recovery. The Long March 10B successfully completed its maiden flight, and its first stage was recovered via a sea-based net. There are several other rockets in development in China that could soon achieve reusability. The country has four land-based spaceports and multiple ocean-going launch platforms and is ready to quickly ramp up its launch cadence.

Try Flow Free → [10] | Download Flow

FROM PROMPT ENGINEERING TO INTENT ENGINEERING

Our thoughts about the smartest way to accomplish a task will become increasingly stupid compared to AI's way of doing it as time goes on. This necessitates a switch from prompt engineering to intent engineering, where you describe the outcome you want rather than how a thing should be done. People should review older prompts and switch their 'how' prompts into 'what' prompts.

IN DEFENSE OF NOT UNDERSTANDING YOUR CODEBASE

People who work on small codebases with low-turnover teams use different methods, practices, and cultures than people who work on large codebases with high-turnover teams. The first group, which says you need to understand the codebase completely, otherwise you can't do good work, is over-represented in online discussions about software engineering. There's nothing wrong with being in a state of partial understanding in many software engineering environments. In large systems, a partial understanding is the best you can do.

THE WAVE HAS ARRIVED”: ZHIPU CO-FOUNDER TANG JIE'S LETTER TO STAFF

Zhipu's co-founder, Tang Jie, has published a full-staff letter announcing a full return to foundation-model research and a two-year plan. The letter covers Tang's views on AGI, AI safety, open source, and more. It also presents a road map to AGI that includes strategic investments in long-horizon tasks, autonomous agent systems, full self-training, and extreme safety governance.

THE REVERSE INFORMATION PARADOX

AI creates a problem where buyers risk giving away knowledge just to use what they bought. They essentially pay for intelligence twice: once with money, and again with the proprietary knowledge they must reveal to make that intelligence useful. Over time, the seller learns more and more about their customers while buyers learn very little about what the seller is learning in return. This reverse information paradox needs to be confronted, as companies should be able to use models without giving up knowledge that makes them unique.

Wispr Flow cleans it up as you go. Try free →

AI 2040 AND THE CULT OF INTELLIGENCE

Current predictions about AI are not grounded in reality.

WHILE NEURALINK DRILLS INTO SKULLS, CHINA'S BRAINCO IS BETTING BRAIN TECH WILL BE SOMETHING YOU WEAR

Hangzhou-based BrainCo makes headbands and caps that pick up electrical signals through the scalp with no surgery required.

WASHINGTON'S BET ON INTEL IS STARTING TO PAY OFF

pay attention. Learn more about sponsorship opportunities.

MEMETIC TRANSFER

A meme is anything that is able to capture cultural transport.

PEPSI CHALLENGE FOR LLMS

LLMs are now good enough that we need new benchmarks to explain their differences.

YOUR BROWSER DOES MATH DIFFERENTLY ON EVERY OS, AND ANTI-BOT SYSTEMS READ THE BITS

The rounding of a cosine can betray the OS that a browser is running on.

Get the Report

CLAUDE CODE ON DESKTOP NOW HAS AN IN-APP BROWSER

Claude Code can now pull up docs, designs, and websites. It can read, click through, and interact with these sites the same way it does with local dev servers. The in-app browser is sandboxed and configurable. Users can choose whether sessions persist. A video demonstration is available in the post.

APPLE SUED OPENAI

Apple accused OpenAI and former Apple executives of stealing confidential hardware information to support OpenAI's device ambitions. The lawsuit alleged that senior leaders directed employees to bypass security procedures and transfer sensitive product details.

ANTHROPIC EXTENDED CLAUDE FABLE 5 ACCESS AGAIN

Anthropic extended Claude Fable 5 access and elevated Claude Code limits through July 19, while OpenAI temporarily removed usage caps for GPT-5.6 Sol. The contrast highlights ongoing uncertainty around Anthropic's long-term model availability.

OWN YOUR WEIGHTS

Owning their own weights can be too complex for most businesses. There needs to be a service where companies can bring their workloads to and get back a task-specific model built on their data, tuned to their tasks, running on infrastructure they control. This would provide the benefits of owning their own weights without having to maintain their own models. It would also allow them to continually build and fine-tune the weights and improve their models.

THE FUTURE WORTH BUILDING IS HUMAN

Thinking Machines aims to develop AI that extends human judgment through customizable, interactive models. It seeks to enable diverse, localized AI adaptations by developing tools, interfaces, and research that allow organizations to align AI with their unique knowledge and values.

THE FUTURE WORTH BUILDING IS HUMAN

Thinking Machines believes the future should be shaped by human knowledge, guided by human will, and decided by human judgment.

to the business. This report is a benchmark you can build on.

Shade from the hype. A path to making AI deliver.

OPENWIKI BRAINS: PROACTIVE MEMORY FOR AI AGENTS

LangChain's OpenWiki Brains introduces proactive memory for AI agents, enabling them to autonomously gather and update relevant context from connected sources like Gmail, Notion, and Twitter. The 0.1.0 version expands beyond code documentation, creating a general-purpose "Personal Brain" through connectors accessing diverse information repositories. By maintaining local wikis, agents can provide consistent context without manual updates, enhancing efficiency across various workflows.

LLM-AS-A-VERIFIER: A GENERAL-PURPOSE VERIFICATION FRAMEWORK

Researchers explore efficient training methods for large language models (LLMs) using sparsity techniques. They aim to reduce computational costs while maintaining model performance. This approach could make training LLMs more accessible for smaller organizations.

PROACTIVE MEMORY FOR LONG-HORIZON AGENTS

A separate memory agent tracked important state and selectively reminded an action agent when relevant information risked being lost. This approach improves pass rates on Terminal-Bench 2.0 and τ2-Bench without modifying the underlying action model.