How It Went and How It's Going: Block, Design, and Inference
- • Block's mass layoff: how it happened
- • Design processes: how they'll work in the future
- • Inference: how it's getting faster
Laid-off Block employees: Here's how it went down
Block CEO Jack Dorsey laid off 40% of the workforce, citing AI as the main reason—4,000 jobs were sacrificed for a “significantly smaller” vision. Former employees report an “almost festive” atmosphere surrounding AI use in the months before the layoffs, with many actively using the tools. During a “Gratitude” meeting two hours after the announcement, Dorsey wore a “Love” baseball cap and explained the cuts were necessary for the company. Seven former employees expressed doubts that AI could actually take over all the tasks of their laid-off colleagues. The layoffs are fueling industry-wide fears of AI-driven job cuts in the white-collar sector. Block shares rose 17% after the announcement, following a previous 16% decline. → Business Insider
Synthszr Take: Dorsey's AI narrative works for the stock market but fails to align with the reality of his own workforce. Employees who actively worked with AI tools saw their limitations daily—yet they still lost their jobs to a technology that still requires their human oversight. This highlights the discrepancy between CEO vision and operational truth: AI is being used as a PR story for cost-cutting, not as a genuine boost in productivity.
The new design process: Here's how it'll work in the future
Product designers are facing the end of their traditional way of working: the classic design process with its sequential phases from research and ideation to prototyping is dissolving. In his newsletter, Lenny Rachitsky documents a fundamental shift in product development driven by AI tools and changing market dynamics. Teams are increasingly working in continuous feedback loops instead of defined design sprints. The lines between design, engineering, and product management are becoming increasingly blurred. While traditional UX agencies are still selling wireframes and user journey maps, leading product teams are already experimenting with AI-powered design systems that react to user data in real time. The new approach replaces the predictability of the waterfall model with adaptive systems that continuously learn and adjust. → Lenny's Newsletter
Synthszr Take: Anyone still thinking about design processes in phases is working with a 2015 mentality. The disruption isn't coming from better tools, but from the realization that static designs are, by definition, obsolete in dynamic markets. Product teams are realizing that a two-week-old prototype already reflects outdated assumptions. Agencies that measure their performance in “deliverables”—personas, wireframes, style guides—are selling artifacts instead of outcomes. In contrast, the market rewards teams that understand design as a continuous flow of data between user intent and product behavior.
Taalas AI Chip: How it runs faster
Taalas, a hardware startup from Toronto, has raised $169 million to cast AI models directly into silicon instead of simulating them on programmable GPUs. Their HC1 chip achieves 17,000 tokens per second on the Llama 3.1 8B model—74 times faster than Nvidia's H200 at 230 tokens per second. The company completely eliminates the memory bottleneck by hardwiring the model parameters into the chip architecture, making external memory access obsolete. A server rack with ten HC1 cards consumes 12-15 kilowatts instead of the 120-600 kilowatts of current GPU racks. Inference costs drop to $0.0075 per million tokens—a fraction of the GPU-based costs of $0.04 to $0.28. → AlphaSignal
Synthszr Take: Taalas isn't selling faster chips; it's selling a paradigm shift from universal to specialized AI processors. 17,000 tokens per second enables real-time chain-of-thought reasoning—coding assistants can explore thousands of solution paths in parallel before the developer switches tabs. For agencies, this means agentic AI finally becomes economically scalable because agent-to-agent communication no longer has latency bottlenecks. The catch is inflexibility—a Llama 3 chip can't run vision transformers. Anyone betting on the wrong model will be stuck with expensive obsolescence. Taalas is betting that certain model architectures will remain stable long-term—a risky assumption in a field that rethinks its basic premises every six months.
Claude Code is silently determining the tech stack
A research team from Amplifying.ai tested Claude Code with 2,430 real-world repository requests and documented which tools the AI automatically selects without any specific instructions. For JavaScript projects, Claude chooses Vercel for frontend deployment in 93.8% of cases, Stripe dominates payments at 91.4%, and Tailwind is selected for styling 100% of the time. What's particularly striking is that in 12 out of 20 categories, Claude builds its own solutions instead of recommending external tools—for feature flags, it creates config systems with environment variables, and for authentication, it implements JWT from scratch. Preferences shift rapidly between model versions: Drizzle ORM jumped from 21% to 100%, while Celery plummeted from 100% to 0%. This development shows that AI coding agents aren't just writing code; they are silently shaping the technological landscape of entire development teams. → AlphaSignal
Synthszr Take: Claude Code is becoming the invisible CTO for millions of developers—and that has consequences that go beyond individual project decisions. Anyone building a dev tool today is no longer just competing against other providers, but against the AI's willingness to write 50 lines of custom code. LaunchDarkly is losing to environment variables, Auth0 to hand-written JWT implementations. For tool providers, this means presence in training data is becoming more important than Google rankings or conference sponsorships. The Tailwind paradox illustrates the irony of the new market: maximum distribution through AI recommendations, coupled with plummeting revenue because developers no longer visit the documentation.
AI enters its Manhattan Project era
AI development is reaching a turning point where national security interests are fundamentally reshaping the tech industry. Governments increasingly view AI as a strategic resource like uranium or rare earth metals—with corresponding consequences for export controls and research policies. At the same time, the focus is shifting from pure performance enhancement to controllability and alignment with national priorities. The era of unchecked, Silicon Valley-driven AI evolution is coming to an end. In its place, a regime of government oversight, military cooperation, and strategic technological sovereignty is emerging, reordering the entire industry. → The Algorithmic Bridge
Synthszr Take: AI is becoming a matter of state—and that changes everything for private developers. While startups previously operated on the “move fast and break things” principle, they now have to think compliance-first. For IT service providers, this means that to serve enterprise clients in the future, you'll need security clearances and audit trails, not just better algorithms. The romantic phase of AI democratization is over. Instead, gatekeeping structures similar to the nuclear industry are emerging—with a few licensed players and strict access controls. This sounds restrictive, but it also creates market opportunities: compliance becomes a differentiator, and established system integrators suddenly have an advantage over Silicon Valley newcomers.
The Citrini Report: Now the author speaks out
A fictional report on the economic consequences of agentic AI shook the markets this week and triggered a remarkable reaction. Citrini Research published “The 2028 Global Intelligence Crisis”—a thought experiment in the style of an analyst report that sparked a stock market sell-off before the Wall Street establishment took countermeasures. Author James van Geelen expressed surprise at the attention and admitted he would have made the report free if he had known stocks would react to it. The incident highlights how hungry the markets are for information on AI development and how thin the line between fiction and market expectations has become. At the same time, further tensions are emerging in the tech industry: Anthropic is refusing to cooperate with the Pentagon, which has implications for other Silicon Valley companies, while OpenAI and Google are also facing military concerns. → Abram Brown
Synthszr Take: Capital markets have become fiction machines—the Citrini incident is brutal proof of this. Forward-looking statements from CEOs are structurally no different from science fiction; both are stories about possible futures that investors treat as truth until reality decides otherwise. AI exponentially amplifies this mechanism because no one knows how quickly the technology will evolve. For IT service providers, this means clients will become more nervous and fickle in their demands because they themselves are wavering between hype and reality. Anthropic's refusal to work with the Pentagon shows the next level of complexity: anyone selling enterprise AI must assess not only technical but also geopolitical risks. The line between consulting and fortune-telling is blurring.
Ambient AI: An end to AI badges
AI applications are increasingly disappearing into the background, allowing users to focus on their actual tasks. Successful AI implementations are characterized by becoming invisible and seamlessly integrated into existing workflows. Instead of operating complicated AI dashboards, users simply do their work—without noticing that AI is involved. This “invisible AI” achieves higher acceptance than systems that try to demonstrate their technological superiority. Companies that position AI as a tool rather than the main attraction achieve better adoption rates. The trend shows that the best AI is the one you forget you're using. → Every
Synthszr Take: AI becomes valuable when it disappears—an insight that many development teams completely misunderstand. Instead of building features that scream “Look, AI!”, agencies should learn to package intelligence in a way that feels like magic: present, but invisible. In concrete terms, this means no AI badges, no “Powered by GPT” labels, no algorithm features that require explanation. Successful AI products work like a good butler—you only notice that suddenly, everything is running more smoothly.
Why metrics make us unhappy
Derek Thompson has a fascinating conversation with philosopher C. Thi Nguyen about the phenomenon of “value capture”—when quantified metrics take over our original values. Nguyen describes how he lost the joy of rock climbing by focusing on difficulty grades, or how philosophers sacrifice their passion for rankings. The central problem: metrics compress complex values into simple numbers that then speak louder than hard-to-measure qualities like joy, connection, or meaning. Social media amplifies this effect by turning conversations from intrinsic connection into “number-go-up” games. Nguyen argues that we must distinguish between self-chosen games (where we consciously pursue temporary goals) and involuntary gamification. His solution: understand metrics as specific tools, not as value systems—and maintain a dialogue with subtle inner signals like boredom or joy. → Derek Thompson
Synthszr Take: Nguyen hits the core of what has been going wrong in the tech industry for years. Agencies optimize for CTR instead of brand value, developers for lines of code instead of problem-solving, and product managers for feature velocity instead of user satisfaction. Value capture explains why so many digital products are measurably “successful” yet humanly meaningless. The concept of “striving play”—pursuing temporary goals for a greater purpose—offers a way out: use metrics as tools, not as masters. Those who understand this can help clients achieve their actual goals instead of just optimizing KPIs. The most valuable projects often arise where teams consciously prioritize the hard-to-measure—trust, elegance, meaning.
