Bilingual

The gauge broke: devs felt 20% faster with AI, measured 19% slower
24 days ago
- A controlled trial found experienced developers felt 20% faster but were actually 19% slower when using AI tools in familiar codebases.
- AI speeds up typing, which isn't the bottleneck for experts, and adds overhead like prompting, waiting, and reviewing often subtly wrong outputs.
- Team-level data shows increased pull request volume and size but no net delivery gain, with more code churn and less refactoring, indicating verification has become the new bottleneck.
- Tool builders are shifting focus to agent-first IDEs, emphasizing verification over generation, aligning with the identified bottleneck in development workflows.
- The feeling of speed is misleading; successful teams must measure actual production outcomes and restaff review stages, rather than relying on perceived velocity.
AI saves about 3% of your hours, and almost none of it reaches the money
22 days ago
- AI saves about 2.8% of work hours in real-world settings, but these time savings rarely translate into higher earnings or profits.
- Task-specific AI gains are substantial (e.g., 40% faster writing, 14% more issues resolved per hour), but they shrink when applied across entire jobs due to tasks AI can't handle and verification overhead.
- AI's competence is 'jagged'; it performs well on structured tasks like writing and customer support but fails confidently on tasks outside its range, leading to costly errors.
- Only 3-7% of productivity gains from AI reach workers' pay, with most organizations seeing zero return on investment despite significant spending.
- To capture AI's value, focus it on high-volume, repeatable tasks it excels at, avoid using it outside its strengths, and deliberately convert saved time into revenue or cost savings.
Apollo economist says a 'painful repricing' of AI markets is possible
19 days ago
- Top economist warns AI productivity gains are concentrated in tech sector, with slow deployment and ROI lagging for most Fortune 500 companies.
- Regulatory hurdles, data protection, and workflow integration delay AI adoption; current market pricing may outpace actual earnings, risking a painful repricing.
- Profit margins show disparity: Magnificent Seven increased from 15% to 25% (2023-2026), while S&P 493 remained around 10%, highlighting limited broader impact.
- MIT study found only 5% of companies saw meaningful ROI from generative AI pilots; companies may slow AI spending if ROI is not quickly realized.
- Ford hired veteran engineers to train staff and reprogram AI tools, acknowledging human oversight is essential for effective AI implementation.
- AI costs currently exceed human labor costs; token optimization efforts indicate companies struggle to get value from AI investments.
- AI shame and pressure to adopt lead to ineffective deployment without clear use cases; BCG study shows saved time from AI often not used strategically.
- Ricoh case study: Outsourcing insurance claims to AI tripled costs vs. manual work, with modest headcount reduction (44 to 39), underscoring high initial time and money investments.
Claude Cowork is coming to mobile and web
18 days ago
- Claude Cowork is expanding to mobile and web, allowing sessions and files to sync across devices, with beta access rolling out starting with Max users.
- Cowork enables Claude to handle tasks across files, calendar, email, messaging, and other tools, with over 90% of usage involving non-software development tasks like business operations and content creation.
- Key changes include work following users across devices, background task continuation even when devices are offline, and user control over decisions, with questions sent to the user's phone when needed.
- Desktop remains the primary platform for deep work, while web and mobile access broadens availability, with chat and Cowork now integrated into a single home interface.
- Users can start Cowork sessions from the Claude website or mobile app sidebar, with extended usage limits through August 5 to encourage trying larger tasks.
Why organizing knowledge in the age of AI sucks, and how I solve it
11 days ago
- The author spends significant time in AI chat sessions for personal and professional tasks, but faces frustrations in retrieving and reusing context from past conversations.
- Current tools like Claude Projects, linear issues, or doc platforms fail to align with human mental models for organizing knowledge, leading to inefficiencies in storing and loading AI context.
- Key insights include the mismatch between session-based organization and human memory, the lack of meta-context in AI readings, and the overlap between human and AI memory needs for natural language artifacts.
- The author built a minimal solution using Claude Code skills to create a knowledge base with folders as topics and files as items, enabling better context capture and retrieval.
- Future plans involve developing an Obsidian-like view layer and more granular control over context loading to enhance the system.
OpenAI's first branded hardware is a light-up keyboard?
10 days ago
- OpenAI launches its first branded hardware device, the $230 Codex Micro, a specialized mini-keyboard.
- It features RGB-lit frosted keys that provide color-coded live feedback on up to six Codex threads (idle: white, thinking: blue, complete: green, needing feedback: amber, error: red).
- A tap on a lit key brings the relevant Codex window on-screen, simplifying thread monitoring for desktop users.
- The device includes six default buttons for common Codex tasks (e.g., accepting changes, push-to-talk), but functions are remappable via software and customizable with included keycaps.
- Beyond Codex-specific use, it supports five additional programmable function layers for general computing shortcuts.
- It's a limited-run collaboration with Work Louder and targets users who monitor AI agents, though mobile apps may offer more detailed remote monitoring.
The AI Productivity Illusion
9 hours ago
- Generative AI hasn't improved productivity at the company or economy level despite enabling faster task completion, as true economic productivity depends on output quality and input costs, not just speed.
- The AI infrastructure build-out already costs $1.5 trillion, and generative AI's error-proneness—like fabrications or errors in code/medicine—reduces output value and increases quality control costs.
- AI's economic value is limited because demand for its outputs isn't as high as transformative technologies like the Industrial Revolution, and it may harm productivity in fields like cybersecurity by enabling attacks.
- The apparent 'productivity paradox' may be an illusion: generative AI's gains from speed are offset by value destruction from unreliability, leading to no net economic improvement.
The Many Claudes of My Code
5 hours ago
- The author initially doubted AI's coding abilities, treating it like Stack Overflow and manually reviewing code to maintain personal style.
- Claude Code demonstrated AI's ability to handle maintenance tasks and reduce the need to learn specific frameworks like Laravel.
- Newer AI models (e.g., Fable, GPT 5.6-Sol) now outperform the author in 80% of coding tasks, leading to acceptance and excitement about productivity gains.
- AI's fallibility is mitigated by deterministic tools that validate code, enabling self-checking with speed and sustained attention.
- The author previously delegated grunt work to a human remote developer (Hercules), which helped reframe AI as a team member rather than a threat.
- The author uses fast, less reasoning-heavy AI models (like Sonnet, 5.6-Luna) as 'Coders' for boilerplate and low-creativity tasks.
- Sophisticated AI models (like Fable/Sol, Opus/Terra) act as 'Reviewers' in adversarial testing, finding bugs and passing issues back to Coders until resolved.
- An 'Assistant/Orchestrator' (e.g., Codex CLI with 5.6-Sol) delegates maintenance, research, and coding tasks, persisting for large, one-shot tasks.
- More AI agents can lead to hallucinations, review loops, and wasted tokens; balance is key, and professional developers must read and understand code to avoid responsibility for failures.
- Skill atrophy is a real risk; vibe coding without deep understanding can lead to losing industry relevance, so AI should accelerate existing skills rather than replace them.
- Controlling the project structure and building with intention makes AI more effective, while aimless use wastes money.
- The author emphasizes that AI is a tool to accelerate human work, not replace human judgment and responsibility.

Hasty Briefsbeta

ai productivity

The gauge broke: devs felt 20% faster with AI, measured 19% slower

AI saves about 3% of your hours, and almost none of it reaches the money

Apollo economist says a 'painful repricing' of AI markets is possible

Claude Cowork is coming to mobile and web

Why organizing knowledge in the age of AI sucks, and how I solve it

OpenAI's first branded hardware is a light-up keyboard?

The AI Productivity Illusion

The Many Claudes of My Code