AI Reshapes Work and Discovery

The productivity puzzle appears to be solving itself. US productivity growth hit 2.7% in 2025, nearly double the decade-long average, while AI-exposed sectors simultaneously cooled their entry-level hiring. This isn't a coincidence. It's the signature of what economists call "skill-biased technical change" at work, where existing workers become more productive rather than companies scaling headcount.

The implications ripple outward. Alibaba's latest model adds visual agentic capabilities, pushing AI from assistive tool to autonomous executor. Meanwhile, researchers deploy these systems for antibiotic discovery, targeting problems where traditional R&D economics failed. These applications share a common thread: AI's value emerges not from replacing human judgment wholesale, but from augmenting domain expertise in specific, high-value tasks.

Yet the structural question looms. If companies can achieve substantial productivity gains without proportional hiring, what happens to the traditional career ladder? Entry-level positions have historically served as training grounds, converting fresh talent into experienced practitioners. When that pipeline narrows, even as productivity climbs, we create a paradox: short-term efficiency gains that may undermine long-term workforce development. The data suggests we're not just automating tasks. We're restructuring how organizations build and deploy human capital.

Deep Dive

AI Finally Makes Economic Sense for Hard Science Problems

The antibiotic discovery work from César de la Fuente's lab at Penn represents something more significant than another AI research win: it solves the economic problem that has plagued drug discovery for decades. Traditional antibiotic development costs run into billions, with timelines stretching 10-15 years and high failure rates. The return on investment has been so poor that most pharmaceutical companies abandoned the space entirely, even as antimicrobial resistance threatens to kill 8 million people annually by 2050.

What makes this work noteworthy is the inversion of traditional R&D economics. De la Fuente's team has compiled a library of over one million genetic recipes for antimicrobial peptides by training AI models to recognize functional sequences in everything from extinct species (woolly mammoths, giant sloths) to archaea and venom. The compute cost to screen these candidates is marginal compared to wet lab work. More importantly, the approach shifts from brute-force mechanical screening (digging through soil samples, testing compounds randomly) to targeted pattern recognition across massive biological datasets. This is the J-curve that matters for founders: AI dramatically compresses the time and cost of the exploration phase while increasing shot quality.

For biotech investors, the implication is clear: AI's value in drug discovery isn't replacing medicinal chemists or clinical trials. It's making previously uneconomic research questions tractable. The antimicrobial resistance market was too risky and unrewarding to attract capital under traditional methods. Now, teams can generate viable candidates for pennies on the dollar, potentially opening entire therapeutic categories that weren't commercially viable before. The work remains early (these peptides haven't reached clinical trials), but the unit economics have fundamentally shifted. That matters more than any individual compound.

India's AI Summit Signals Shifting Capital Allocation Patterns

The India AI Impact Summit this week drew Sam Altman, Dario Amodei, Sundar Pichai, and major capital commitments, but the real story is what happens after the speeches end. India now represents ChatGPT's second-largest user base globally with over 100 million weekly actives, yet the country has captured minimal AI infrastructure investment relative to usage. That gap is closing rapidly, and it will reshape where AI compute capacity gets built over the next 24 months.

Consider the concrete deployments: Blackstone invested $600 million in Neysa for GPU infrastructure, with another $600 million in debt planned to deploy 20,000 GPUs. AMD partnered with Tata Consultancy Services on rack-scale AI infrastructure. Anthropic opened its first office in Bengaluru. India earmarked $1.1 billion for a state-backed VC fund targeting AI and advanced manufacturing. These aren't MoUs or photo ops. This is capital flowing toward physical infrastructure in a market where power, real estate, and engineering talent costs run substantially below US equivalents.

For VCs and founders, the strategic question isn't whether India will build AI capacity, but how that capacity reorients global AI economics. If India deploys meaningful inference infrastructure at 40-60% lower operating costs than US facilities, it doesn't just serve domestic demand. It becomes a competitive inference hub for global applications, particularly as models commoditize and margin pressure intensifies. The more concerning signal for US tech workers: Vinod Khosla's comment that 250 million young Indians should be "selling AI-based products and services to the rest of the world" isn't aspirational. It's a roadmap for how global AI labor markets restructure when capability equalizes but cost structures don't. The BPO playbook worked once. It will work again.

Signal Shots

Anthropic Collapses Claude Code Output, Developers Push Back : Anthropic updated Claude Code to hide file names and details during AI coding sessions, collapsing output to summaries like "Read 3 files" instead of showing which files the AI accessed. Developers immediately objected, citing security concerns, context verification needs, and token cost management. Developers want to catch mistakes early and audit AI behavior, particularly when working on complex codebases where pulling wrong context wastes expensive API calls. Anthropic's Boris Cherny suggested developers "try it for a few days," then partially walked back the change by repurposing verbose mode. Watch whether this becomes a pattern across AI coding tools as they hide reasoning to simplify interfaces, and whether that opacity creates new categories of costly errors.

GPT-5 Follows Law Better Than Human Judges, Raising New Questions : University of Chicago researchers tested OpenAI's GPT-5 against 61 federal judges on state law conflicts questions and found the model applied correct legal doctrine 100% of the time versus 52% for judges. Google's Gemini 3 Pro also scored perfectly, while other models ranged from 50% to 92% compliance. The gap exists because judges exercise discretion when doctrine allows interpretation, which the researchers argue is actually a strength rather than weakness in complex cases. Watch how legal profession and policymakers respond to perfect formalism from AI systems, particularly whether anyone wants machines dispensing justice when following rules to the letter produces morally questionable outcomes.

Defense Department Threatens to Cancel Anthropic Contract Over Usage Limits : The Pentagon is demanding AI companies allow "all lawful purposes" for military use, but Anthropic is resisting, according to Axios, with Defense officials threatening to pull a $200 million contract in response. The standoff centers on Anthropic's hard limits around fully autonomous weapons and mass domestic surveillance, not individual operations. This matters because it forces transparency on exactly which capabilities AI labs will and won't enable for government customers, creating precedent other companies will have to navigate. Watch whether other frontier labs hold similar lines or quietly accommodate Pentagon demands, and how that affects their positioning with enterprise customers concerned about dual-use applications.

Indian Startup Attacks Datacenter Power Conversion Waste : Bengaluru-based C2i Semiconductors raised $15 million from Peak XV to build integrated "grid-to-GPU" power systems that aim to cut energy losses by 10% compared to current datacenter designs, which waste 15-20% of electricity during voltage conversion. The approach treats power delivery as a single system rather than discrete components, potentially saving 100 kilowatts per megawatt consumed with downstream benefits for cooling and GPU utilization. Power is becoming the binding constraint on AI scaling faster than compute availability, making even incremental efficiency gains economically significant at hyperscale. Watch whether C2i's silicon validates the claims when it returns from fabrication in April-June, and whether power delivery becomes as hotly contested as GPU supply chains over the next 24 months.

California Billionaires Pour Tens of Millions Into State Politics : Tech billionaires are spending heavily to influence California politics as Governor Gavin Newsom reaches term limits, with major campaigns targeting a proposed 5% billionaire tax, AI-friendly candidates, and new Super PACs. Peter Thiel donated $3 million to fight the wealth tax while moving to Florida. Meta launched two Super PACs focused on AI regulation with $65 million combined, and crypto executives opened Grow California with $10 million. This represents a significant escalation from tech's traditionally narrow state lobbying focused on specific issues. Watch whether the industry successfully installs a tech-friendly successor to Newsom and blocks the billionaire tax, setting template for how AI boom winners deploy political influence as regulatory pressure builds.

Scanning the Wire

UK Prime Minister sets 'months' timeline for social media age limits : Keir Starmer signals imminent crackdown on platforms with stricter controls for VPNs and AI chatbots as part of child safety push, setting up potential confrontation with Big Tech. (The Register)

Cisco building proprietary hypervisor as VMware alternative : The networking giant is developing its own virtualization layer specifically for Unified Communications Manager and calling applications, offering users a path away from Broadcom's VMware. (The Register)

EU bans destruction of unsold clothing and accessories : New regulations prohibit apparel companies from destroying excess inventory, forcing brands to find alternative disposal or redistribution methods for unsold goods. (Hacker News)

Ring kills neighborhood camera sharing feature after Super Bowl ad backlash : Amazon's Ring discontinued its partnership with Flock Safety for the "Search Party" feature following privacy concerns triggered by commercial showing lost dog reunited via connected cameras. (New York Times)

Prediction market Kalshi generates $1.3 billion in sports betting revenue : Sports wagers on the platform now pose material threat to DraftKings and Flutter, driving stock declines as investors reassess prediction markets' competitive position against traditional operators. (Financial Times)

NASA demonstrates AI-generated waypoints for Mars Perseverance rover : Agency used Claude AI to analyze orbital imagery and plan 456-meter route across two days, marking shift toward greater autonomy as signal delays make human-controlled driving increasingly limiting. (IEEE Spectrum)

African defense startup Terra Industries raises additional $22 million : The Gen Z-founded company secured follow-on funding within a month of previous round as African governments accelerate domestic defense technology procurement. (TechCrunch)

Blackstone backs Indian GPU provider Neysa with up to $1.2 billion : Investment targets deployment of over 20,000 GPUs as India races to build domestic AI infrastructure capable of serving both local demand and potentially global inference workloads. (TechCrunch)

Fractal Analytics shares fall in muted India IPO debut : First Indian AI company to go public underwhelms investors as enthusiasm for technology collides with broader selloff in software stocks and questions about near-term profitability. (TechCrunch)

MicroVision targets sub-$200 solid-state lidar for automotive market : Company claims design could drop below half current pricing with long-term goal of $100 per unit, potentially making 3D sensing economically viable for advanced driver assistance systems beyond premium vehicles. (IEEE Spectrum)

Outlier

Dating Apps Hit 3.5 Million Downloads in Saudi Arabia : Dating app adoption in Saudi Arabia has grown every year for five consecutive years, reaching 3.5 million downloads in 2025 as social liberalization enables more casual relationships. This isn't just about romance. It's a signal of how digital infrastructure can outpace cultural change, creating new social behaviors before institutions catch up. When technology enables coordination around previously taboo activities, it doesn't just reflect social change. It accelerates it by making the previously invisible visible and the previously isolated connected. Watch how other Gulf states respond as apps create facts on the ground faster than traditional gatekeepers can regulate them, and whether this pattern repeats across other culturally sensitive categories from financial services to education.

The productivity gains are real, the questions are harder, and somewhere a judge just got outscored by a machine that can't tell right from wrong but follows every rule perfectly. That should keep us all appropriately uncomfortable.