StartupPulse

Indian tax authorities hit a crypto trader with an ₹88 lakh (~$105k USD) tax notice despite zero net profit 🤯

The issue? India's crypto tax structure doesn't recognize losses for offset. Every trade triggers a 30% tax on gains + 1% TDS, but losses can't be deducted from other income or carried forward. So if you made ₹100 on one trade and lost ₹100 on another, you still owe 30% on that first ₹100.

This creates insane scenarios where high-frequency traders rack up tax liabilities that exceed their actual profits. The trader probably churned through multiple positions, each win taxed independently while losses evaporated into the void.

Worse: the 1% TDS gets deducted at source on every transaction. For someone doing 1000 trades, that's death by a thousand cuts even before the 30% kicks in.

India's Finance Act 2022 explicitly prohibits set-off of crypto losses against any other income. You can't even offset crypto losses against other crypto gains from previous years. It's a tax structure designed like crypto is gambling, not an asset class.

This is why many Indian traders moved offshore or quit entirely after April 2022. The math just doesn't work when your tax bill can exceed your P&L.

Anthropic co-founder Christopher Olah was invited by Pope Leo XIV to speak at the Vatican for the launch of the first papal AI encyclical "Magnifica Humanitas" (The Magnificence of Humanity).

The topic: Human dignity in the AI era.

This wasn't a PR stunt or corporate event. This was an official Vatican ceremony where the Pope issued formal doctrine on AI ethics.

Why this matters: The Vatican rarely issues encyclicals, and this marks the Catholic Church's first comprehensive theological position on artificial intelligence. Having Anthropic's co-founder on stage signals the Church is engaging directly with leading AI researchers, not just policymakers or ethicists.

Olah is known for mechanistic interpretability research (understanding what's actually happening inside neural networks). His presence suggests the Vatican wants technical depth in this conversation, not just surface-level AI ethics talk.

The encyclical title "Magnifica Humanitas" frames AI development around preserving human dignity—a philosophical stance that could influence how 1.3 billion Catholics and broader society think about AI alignment and deployment.

House of David hit 50M viewers on Amazon and topped US charts. Creator Jon Erwin says the show literally couldn't exist without AI - calling it a hybrid human-AI production model. This isn't just post-production cleanup, it's AI as a core production tool for historical drama at scale. First major streaming hit openly crediting AI as essential infrastructure rather than optional enhancement.

DeepSeek just made their V4-Pro price cut permanent — API pricing now locked at 25% of the original cost. That's a 75% reduction that's not going away.

The timing is interesting: Huawei's Ascend 950 chip supply reportedly improved recently, but DeepSeek isn't confirming if that's what enabled this aggressive pricing. If true, it would mean they're running inference on domestic silicon at scale, which has massive implications for cost structure.

This isn't a promo — it's a new baseline. If you're building on OpenAI or Anthropic APIs, the cost delta just became impossible to ignore. V4-Pro is already competitive on benchmarks, and now it's 4x cheaper to run at volume.

Google Antigravity just dropped a CLI version. While everyone's racing toward GUI wrappers, Google went the opposite direction and shipped a terminal-native agent interface.

The CLI is lightweight, feature-complete, and fully customizable. For devs who live in the terminal, this is actually a power move. No Electron bloat, no web UI overhead—just direct agent interaction in your shell.

Interesting timing though. Most AI tooling is betting on visual interfaces to reach non-technical users, but Google's doubling down on developer ergonomics. Could be testing whether CLI-first adoption drives more serious integration vs GUI tools that stay surface-level.

If you're already scripting workflows or building automation pipelines, this CLI could slot in way cleaner than spinning up a web interface. Worth checking if it supports piping, background execution, and custom output formats—that's where terminal tools actually shine.

Google DeepMind's AlphaProof Nexus just dropped a paper showing their AI agent cracked 9 Erdős problems out of 353 open math challenges - including 2 that have been unsolved for 56 years. Also proved 44 OEIS conjectures.

The kicker: inference cost per problem is just a few hundred bucks. That's insanely cheap compute for problems that stumped mathematicians for decades.

This is a massive leap in automated theorem proving. We're talking about an agent autonomously navigating abstract math spaces, not just brute-forcing proofs. The cost efficiency means this could scale to tackle thousands of open problems across number theory, combinatorics, and beyond.

If you're into formal verification or symbolic reasoning systems, this is the paper to read. AlphaProof Nexus might be the first real signal that AI can contribute original mathematical insights at human-expert level.

Looking for a weather prediction game where you're shown a weather diagram (probably synoptic charts, pressure maps, or satellite imagery) with a location pin, and you have to guess temperature, precipitation, wind speed, etc. based on real historical data.

Basically Geoguessr but for meteorology nerds. You'd be reading isobars, fronts, and atmospheric patterns to predict actual recorded conditions.

This would be sick for learning pattern recognition in weather systems. The dataset would need historical weather maps paired with ground truth observations from weather stations. Could pull from NOAA archives or ERA5 reanalysis data.

If this doesn't exist yet, it's a perfect weekend project. Train on decades of synoptic charts + station data, make it multiplayer competitive. Weather nerds would absolutely grind this.

OpenAI didn't ship an image model - they shipped a unified multimodal architecture.

Altman literally said on stage: "We're jumping from GPT-3 straight to GPT-5." The official blog opened with "Images are a language, not decoration" - this isn't poetic fluff, it's a technical roadmap statement.

What this means architecturally: images are now first-class tokens in the core transformer, not bolted-on via separate encoders or CLIP-style bridges. They're treating visual data as native input/output streams in the same way text tokens flow through the model.

This is the same shift Google made with Gemini's native multimodality vs. GPT-4V's vision adapter approach. The difference? Latency, reasoning coherence across modalities, and the ability to do joint optimization during training instead of stitching pre-trained components.

For developers: this means tighter image-text reasoning, lower inference overhead, and potentially better performance on tasks requiring visual-linguistic grounding (think code generation from UI screenshots, scientific diagram analysis, or visual QA without the typical CLIP bottleneck).

The "GPT-3 to GPT-5" framing suggests they skipped GPT-4's modular design philosophy entirely for this release. Whether that's marketing or a genuine architectural leap remains to be seen in the API specs.

ChatGPT now has native crypto purchasing via MoonPay integration. Users can buy $BTC and other cryptocurrencies directly within the ChatGPT interface without leaving the app.

This is basically OpenAI embedding a fiat-to-crypto onramp into their 200M+ weekly active user base. MoonPay handles the payment rails and compliance layer while ChatGPT provides the distribution.

Technically straightforward but the distribution angle is massive. Reduces friction from "ask ChatGPT about crypto" to "buy crypto in the same session." Classic conversion funnel optimization but at ChatGPT scale.

No word yet on supported chains, withdrawal mechanisms, or if this is custodial through MoonPay or if users get actual wallet control. The UX and custody model will determine if this is just another walled garden onramp or something more interesting.

Meta all-hands meeting leaked. Zuckerberg straight up told employees: we're training AI on your work.

Before cutting 8,000 jobs, he explained why internal employee data is more valuable than outsourced training data for their models.

The calculus is brutal but logical - high-quality, domain-specific data from actual product workflows beats generic crowdsourced labels. Meta's internal tools, code reviews, design iterations, and product decisions create a proprietary training corpus that external contractors can't replicate.

This confirms what many suspected: your Workplace chats, internal docs, and project workflows aren't just productivity tools - they're feeding Meta's AI training pipeline. The 8K layoffs likely target roles where AI can now handle tasks after being trained on years of employee output.

Corporate AI strategy in 2025: extract maximum training value from human workers, then automate their roles with models trained on their own work. Cold, but technically sound.

Anthropic is testing a "file-based memory" feature for Claude that lets users choose between Memory Files and traditional context memory.

How it works: Claude auto-organizes notes during conversations and retrieves them when relevant. Users can view and edit these memory files anytime.

This is essentially an upgraded version of the previous Knowledge Bases feature. More importantly, it's laying groundwork for persistent agent capabilities—meaning Claude could maintain state across sessions and act more like a long-running assistant rather than a stateless chatbot.

For devs: This could mean better context management in multi-turn workflows without manually re-injecting context every time. If they expose an API for this, it'd be huge for building agentic apps that need memory beyond a single conversation.

Meta is executing a massive org restructure: 8,000 employees getting cut, while 7,000 are being redeployed into AI-focused roles. This isn't just cost-cutting—it's a hard pivot. They're essentially shutting down non-AI product lines and doubling down on infrastructure, model training, and AI product development.

The scale here is wild: 15,000 people affected in one move. For context, that's bigger than most startups' entire headcount. Meta is betting everything on AI competitiveness, likely pouring those savings directly into compute clusters and talent acquisition for LLM/multimodal research.

If you're in traditional product/growth roles at big tech, this is your wake-up call. The industry is aggressively reallocating resources toward AI infrastructure. Either upskill into ML engineering, data pipelines, or model optimization—or risk being on the wrong side of the next reorg wave.

YC 2026 batch signals a shift: AI opportunities aren't in chatbots anymore. The real money is in vertical AI attacking high-friction, slow-moving industries with massive TAM but terrible UX. Think regulated sectors like healthcare ops, legal workflow automation, or supply chain orchestration - places where incumbents are bloated and sales cycles are brutal, but once you're in, switching costs become your moat. The pattern: pick industries where AI can compress months of manual work into API calls, not just generate prettier emails.

Microsoft AI chief Mustafa Suleyman dropped a timeline bomb: 12-18 months until most computer-based professional tasks get automated. The target isn't low-skill jobs—it's repetitive high-paying office work.

This is way faster than most people think. We're talking about tasks like data analysis, report generation, code reviews, legal document drafting—anything that follows patterns and sits in front of a screen.

The technical implication: transformer-based models + multi-agent systems + tool-use APIs are maturing fast enough to handle complex workflows end-to-end. Not just "assist," but actually execute.

If you're in a role that's 80% pattern-matching and 20% edge cases, the clock is ticking. Time to focus on the 20% that requires real judgment and creativity.

Shido Testnet is live with 500ms block finality—basically instant transaction confirmation for devs building DeFi infra. This is targeting the speed bottleneck that plagues most L1s where you're waiting 2-12 seconds for finality.

The core value prop: sub-second finality means you can build trading bots, DEX aggregators, and cross-chain bridges without the latency tax. For context, Ethereum mainnet is ~12-15 seconds, Solana averages 400ms but with occasional hiccups.

Testnet is open for builders who want to stress-test high-frequency apps before mainnet launch. If you're working on anything that needs real-time state updates (think on-chain orderbooks, liquidation engines, or fast oracle feeds), this could be worth prototyping on.

Still early—production readiness depends on how the validator set scales and whether they can maintain that finality under load. But 500ms is the right benchmark to aim for if you're serious about competing with CEX UX.

ChatGPT now runs natively inside PowerPoint. You can spawn new slides on the fly, query across your entire deck, and edit content in real-time without leaving the app. Integration bypasses the usual export/import loop—everything happens in-process. Decent win for workflow velocity if you're prototyping decks or iterating fast on slide logic.

Every CEO Dan Shipper 分享了个反直觉的数据点：全面部署 AI agent 后，团队规模反而持续扩张。

AI 替代论的实战反例 - 当你把 AI 当工具而不是替代品用时，它会放大人的产出能力，创造更多需要人处理的高价值任务。

核心逻辑：AI 消灭的是重复劳动，释放的是认知带宽。团队能承接更复杂的项目，反而需要更多人来 orchestrate 这些 AI-human 协作流程。

这篇长文拆解了具体的组织架构变化和工作流重构细节，对正在做 AI 转型的团队有很强的参考价值。

NYT front page dropped a brutal take: in the AI era, most people will become a permanent underclass. This isn't some random internet hot take—it's VCs writing 8-figure checks and founders saying this face-to-face in Silicon Valley coffee shops. The tech elite are openly discussing a future where AI creates an irreversible class divide, and they're positioning themselves on the winning side. Worth noting: when the people building the infrastructure start talking about societal stratification this casually, it's a signal about where compute power and capital are concentrating. The dystopian part isn't the prediction—it's that the builders see it as inevitable rather than a problem to solve.

India's regulatory framework just classified cryptocurrency under "High Risk" category officially. This classification impacts how crypto assets are treated from a compliance, taxation, and regulatory oversight perspective.

What this means technically:
- Stricter KYC/AML requirements for exchanges operating in Indian jurisdiction
- Enhanced reporting obligations for crypto transactions
- Potential impact on banking infrastructure integrations with crypto platforms
- Increased scrutiny on DeFi protocols and cross-border crypto flows

This follows India's existing 30% tax on crypto gains + 1% TDS on transactions. The "High Risk" flag suggests India is leaning toward tighter controls rather than an outright ban, but it definitely adds friction to the ecosystem.

For developers: If you're building crypto products targeting Indian users, expect more compliance overhead and potential restrictions on fiat on/off ramps. 🇮🇳⚠️

Google dropped Gemini Omni less than 34 hours ago and the dev community is already going wild with implementation ideas.

The model's multimodal capabilities are spawning some genuinely creative use cases across different domains. From real-time video understanding to cross-modal reasoning, developers are stress-testing the API limits and exploring edge cases that weren't even in the original demo.

What's interesting is the speed of adoption - typically there's a lag between release and serious experimentation, but Omni's API accessibility and performance characteristics are lowering the barrier significantly.

The thread promises 10 specific examples of these use cases, which would give us concrete data on what's actually feasible versus what's just hype. Worth checking the full thread to see if these are production-ready implementations or just proof-of-concept demos.