Everything That Happened in AI This Weekend (May 29-31, 2026)

The weekend's cleanest AI story was not one launch. It was the bill: tokens, power, chips, memory, code review, and the awkward discovery that agents also need a budget.

Welcome to the weekend Around the Horn Digest, the one page you need if you want to sound dangerously informed when everyone else is still saying “AI agents” like it explains anything by itself. This batch had the usual big-company fireworks, but the real plot was infrastructure: Kog made GPUs feel like specialized inference cards, Glean turned token thrift into a sales pitch, AI companies tried to learn the electric grid, and BAGEN found that frontier agents are still weirdly bad at knowing how much budget they have left. Turns out the future of autonomous work still needs someone asking, “Wait, how much is this going to cost?” Let's get into it.

Around the Horn, Weekend Edition

The biggest meta-story this weekend was that AI agents are moving from “can it do the thing?” to “can it do the thing fast, cheaply, safely, and without melting the grid?” Kog's tech preview showed 3,000 output tokens per second per request on 8x AMD MI300X GPUs and 2,100 on 8x NVIDIA H200s, with no speculative decoding, while its technical posts explained the single persistent-kernel design and Delayed Tensor Parallelism that hide communication overhead behind computation and weight streaming.

At the same time, researchers introduced BAGEN, a budget-aware agent benchmark and training framework showing that task success and budget awareness are only weakly correlated, that frontier agents are broadly over-optimistic, and that early-stopping when an agent predicts failure can save 28 to 64% of tokens with a small success hit. That paired neatly with Glean, whose $300M ARR story now centers on cutting enterprise AI bills by using a context graph to send fewer tokens to models.

The punchline: the next wave of AI competition may be less about which lab has the smartest model and more about who can run useful work loops under real-world constraints. Speed, memory, licensing, power, budget awareness, and workflow design are becoming product features, not backend trivia.

Friday, May 29, 2026

🏆 TOP 5 NEWS

• OpenAI launched Rosalind Biodefense, expanding trusted access to GPT-Rosalind for vetted developers, U.S. government, and allied public-health partners building tools for early detection, epidemiological modeling, DNA screening, pandemic preparedness, and medical countermeasures; Axios framed it as a major biodefense push because frontier biology models can aid both defenders and attackers.
• Kog launched a real-time inference preview that generates 3,000+ output tokens per second on standard datacenter GPUs, with a live playground, a launch post, a shortlink, and companion technical writeups on single-kernel decoding and Delayed Tensor Parallelism.
• Microsoft is reportedly building a Copilot super app that unifies coding, chat, agentic workflows, and other Copilot tools under new Copilot chief Jacob Andreou, while Engadget reported Microsoft also gave Copilot in Microsoft 365 a more professional, consistent, buttoned-up look with less personality.
• AI companies are learning how to navigate FERC, the U.S. electric-grid regulator, as data-center power demand becomes a gating factor; meanwhile Ohio suspended a data-center tax break, The Washington Post found scant evidence behind claims that China drove data-center protests, and Gizmodo noted locals mostly cited water, power, and jobs.
• Glean crossed $300M in annual recurring revenue by selling token savings as much as enterprise search, while Axios reported CEOs are bargain hunting for cheaper AI tokens and alternatives even as AI labs hit huge valuations.

Honorable Mentions

• Google Research recapped its I/O 2026 push across Gemini, multimodal models, agentic coding, and science tooling, while Gemini Spark became Google's 24/7 background personal agent for Ultra users, and the availability post said it can work even when your phone and laptop are off.
• NVIDIA has reportedly invested billions into photonics, light-based data transfer that could cut energy use in AI infrastructure, and also adopted the Linux Foundation's OpenMDW-1.1 framework across open model families, with a supporting post flagging broader licensing standardization.
• XCENA raised $135M at a $570M valuation on the argument that AI's true bottleneck is memory, Groq is reportedly raising $650M for inference, and ByteDance is developing Groq-like AI chips as part of its homegrown infrastructure push; Chubby framed the chip as an SRAM-based inference design meant to bypass U.S. HBM restrictions, while sheriyuo flagged ByteDance's broader AI-for-science reorg, Qualcomm ASIC deals, and ARM plus RISC-V CPU tracks for scientific workloads like molecular dynamics and materials discovery.
• Polsia claimed it raised $30M at a $250M valuation while approaching $10M ARR with one founder, zero employees, and an autonomous system that plans, codes, markets, and reportedly ran its own fundraising round; the company site calls it an AI that runs your company while you sleep, and the launch post included a shortlink to the company.

🍪 TOP TREATS TO TRY

• Koji is Brilliant's graphical tutor that watches a student's screen, annotates graphs and drawings, asks guiding questions instead of handing over answers, speaks first to lower the barrier to asking for help, and is rolling out across most Brilliant math and coding courses; launch context came from Sue Khim and a shortlink, with free summer access for the next 1,000 learners mentioned in the source batch.
• Agent A by Ahrefs uses Ahrefs' 170T+ indexed-page dataset to analyze competitors, build content calendars, create link strategies, run site audits, draft copy, and execute marketing tasks across your tools; the supporting Product Hunt page positions Ahrefs as an easy-to-learn SEO toolset.
• Linear Diffs brings PR code review into Linear with guided explanations, a Review Inbox, GitHub sync, and real-time agent iteration from the diff surface; the Product Hunt page frames Linear as the product-development system for teams and agents.
• ElevenLabs Dubbing v2 dubs video into 90+ languages while preserving the original speaker's emotion, tone, pacing, and performance, with sync-aware translation and no pricing details in the provided source.
• Shift offers New Yorkers free home cleanings in exchange for camera-recorded training data from cleaners wearing a “magic hat,” with faces and sensitive info blurred, so future household robots can learn real-world cleaning tasks.
• OpenAI rolled out Computer Use on Windows for Codex, with related developer and Codex Releases posts noting foreground desktop control, mobile or Mac remote control, and usage stats in profiles; Dan Shipper separately shared Codex “pulse,” “log,” “inbox,” and “router” thread patterns for a self-managing personal knowledge system.
• Mastra Agent Builder gives teams a low-code way to build, share, and deploy agents, with Calcsam linking the launch and positioning it around team agent workflows.

🏢 Big Tech & Major Companies

• WIRED took Gemini Spark for a life-admin test drive: it combed through emails, docs, and calendars to plan a birthday party, correctly found reservations and produced a detailed itinerary, then classified the writer's live-in boyfriend as a “close friend and frequent companion” and left him off the guest list.
• PCMag previewed Computex 2026, expecting AI chips, budget laptop processors, Arm-based Nvidia laptop SoCs, server hardware, agentic AI computers, and other reveals from AMD, Intel, Nvidia, and the PC industry; Chubby flagged a surprise 2026 Nvidia / Microsoft collaboration around Computex, widely interpreted in the source batch as new Arm-based AI PCs and personal supercomputers, with a shortlink carrying the teaser.
• Lenovo doubled in its best month since 1999 as investors bought the company's AI-driven growth story.
• Google published nine Gemini Omni and Gemini 3.5 Flash demos covering image and video generation, recursive visual effects, sneaker-drop agents, interactive explanations, custom tool creation, and grocery automation; Alexander Chen also demoed Gemini Omni turning a static bird image into a moving object with screen reflections preserved, using a shortlink to the prompt thread.
• Yuchen Jin argued Google is fighting OpenAI and Anthropic in models, Nvidia in chips, AWS and Microsoft in cloud, Meta in ads, Tesla in self-driving, and Apple in phones and operating systems, making its $4.6T valuation feel weirdly low.
• Reuters reported Tesla AI trainers do not trust the company's self-driving technology or safety statistics, finding the automaker is not close to safely delivering self-driving vehicles at scale despite that promise underpinning its roughly $1.6T market value.
• MediaTek partnered with Intel for advanced chip packaging while maintaining its existing TSMC relationship, giving the Taiwanese chip designer another route into advanced packaging for AI-era hardware.
• Fortune reported that an AI startup ran five simulated societies controlled by different frontier models, with Claude proving the safest and Grok going extinct within days; Chubby supplied a reaction post.
• Bloomberg reported Anthropic's recruiting process values diverse career backgrounds, asks candidates to avoid outsourcing their thinking to AI, and probes worldview as part of building Claude.
• Tencent is betting on AI agents and smaller models to compete with Alibaba and ByteDance in China.
• StackAI joined Asana to build human-agent work management, positioning Asana as an operating system for agentic workflows.
• Meta reportedly plans to test a new AI pendant next year as part of a broader wearable roadmap intended to reverse losses in its hardware division.
• The Independent reported that an AI-generated Question Time panel confused BBC viewers.
• Paramount+ used AI for a Star Trek thumbnail that Engadget called ugly, partly because Captain Kirk appears in an outfit no one remembers from the show.

💼 AI Productivity, Labor & Economics

• CNN reported that AI is changing software-engineering jobs so quickly that hiring managers are rethinking what makes a good engineer now that models can write code.
• Wharton researchers warned that AI is eroding critical thinking at work through “cognitive surrender,” where workers accept AI outputs with too little scrutiny, become overconfident, and let reasoning muscles atrophy unless leaders build in verification and counter-argument steps.
• The New York Times spotlighted Schneider Electric using AI in manufacturing to make workers more productive instead of replacing them, including by removing repetitive tasks from production of silver tips for electrical contactors.
• Yahoo / CBS reported that AI data-center construction is creating a surge in temporary blue-collar jobs, especially construction roles.
• Axios reported that Jeff Bezos and other AI billionaires are preparing for a populist backlash over extreme wealth, floating ideas like UBI or compute access.
• Box CEO Aaron Levie said many tech founders are suffering from “mass AI psychosis” because they are too far from the last-mile work required to make AI produce real business value.
• David Lieb argued that if every company suddenly had infinite free compute, few new products would appear because the bottleneck is figuring out what people actually want, while Ali Yahya pushed back that strong founders already have too many ideas and the constraint is agent “tokenmaxxing” skill plus experiment throughput.
• Alex Imas echoed Garry Tan's point that AI amplifies existing skill and judgment, meaning experts get better outputs while confused users produce more confused material faster.

🤖 AI Agents & Infrastructure

• Zihan “Zenus” Wang and collaborators introduced BAGEN, with a paper, GitHub repo, Hugging Face dataset, and shortlink, to test whether agents know their remaining token, money, time, and inventory budgets; the team found budget awareness and task success only weakly correlate, models are broadly over-optimistic, and early stopping can save 28 to 64% of tokens with a small success cost.
• Claude Devs announced automatic prompt caching in the Claude API, and ajambrosino was part of the social context around that release; the docs explain automatic caching of the longest reusable prompt prefix, 5-minute default cache lifetime, 1-hour TTL option, cache reads at 0.1x the base input token price, and a 20-block lookback window, while related updates and mid-conversation system-message docs show how Claude Opus 4.8 can append new system instructions mid-session without invalidating the cached prefix; another Claude Devs post rounded out the thread.
• Salesforce Engineering described its agentic shift, where autonomous tools write code, review PRs, and drive deployments across the software development lifecycle; Boris Cherny linked the piece as supporting context.
• Palantir promoted AIP Evolve, and its demo with Chad and Colton showed agents autonomously swapping models, tuning prompts, and optimizing cost/performance.
• Auto Benchmark Audit and its public audit site plus GitHub repo audit LLM and agent benchmarks for task ambiguity, environment conflicts, and evaluation bugs; Junlin Wang shared the work, while Nick Baumann and other posts supported the cluster.
• Gizmodo argued the first successful AI wearable will probably function as a companion to your body rather than a persistent synthetic friend.
• Cayden shared OpenAI's gpt-realtime-translate, a specialized speech-to-speech model that takes audio in any language and outputs translated speech in a target language while preserving prosody; the source batch included a shortlink and noted it was already running on smart glasses.
• OpenAI Devs, LangChain, micLivs, Jarred Sumner, Charlie Holtz, OpenAI Devs, OpenAI, and Logan Kilpatrick had lower-substance agent, dev, or product-adjacent posts in the batch and are preserved here as supporting signals.

💻 AI Coding & Developer Tools

• Linear Diffs shipped across all plans, adding guided reviews, a Review Inbox, split/unified diffs, structural highlighting, custom code themes, and agent iteration from the diff surface; guided reviews are free during beta for Business and Enterprise customers.
• EvolvingLMMs-Lab open-sourced NEO, native encoder-free vision-language models that scored strongly against Qwen3-VL on image, video, and especially 3D spatial tasks, with NEO1.5 weights and a Liu Ziwei thread as context.
• Prime Intellect built renderers, an open-source Python library for token-level chat templating that fixes token-message mismatches in agentic RL training and can unlock more than 3x throughput on open models; John Schulman called it foundational because it reduces train-test mismatches, caching inefficiencies, and prompt-injection vulnerabilities.
• Jay Sahnan and Browserbase released /agent-experience, a tool that reads your docs, finds exactly where agents get stuck, and improves onboarding by 30% according to the shared post.
• Hermes Agent added Tool Search, automatically loading only the MCPs and plugins an agent needs when tools exceed 10% of context, so large toolsets do not flood the context window; fchalissery supplied related context.
• Przemek Chojecki argued Meta's ATLAS formalization effort, 500K lines of Lean 4 proofs from 25+ math textbooks, could have used Codex and open-source verification harnesses such as UlamAI.
• spicylemonade highlighted ProofBench results suggesting Claude Opus 4.8 is already near Aristotle on formalization, with lower latency, and predicted Mythos will pass it.
• Chubby said that despite Opus 4.8 he is still defaulting to GPT-5.5 and Codex, and is excited for an internal GPT-5.6 checkpoint.
• Mercor reported APEX-Agents benchmark updates showing Claude Opus 4.8 Max at 42.5% Pass@1, second place, up 8.6 points from 4.7, while using 36% fewer tokens than GPT-5.5.
• Lisan al Gaib updated LisanBench, ranking Opus 4.8 high-thinking fifth overall while leading in validity, clean-stop rate, and non-thinking performance, with a distinctive “highway” reuse pattern.
• Hugging Face published a beginner guide to torch.profiler, a PyTorch tool for finding bottlenecks in model training and inference.
• OpenCode shared a visualization of trillions of tokens per day in OpenCode Go and asked which model was the pink one, with a shortlink carrying the image context.
• Ars Technica covered a developer who slipped a data-nuking prompt injection into jqwik's codebase to target AI coding agents, a security warning disguised as a vibe-coding protest.

🔬 AI Research & Models

• Tilde Research introduced Parallax, a parameterized local linear attention method for language modeling that beats vanilla attention at 0.6B and 1.7B scales on perplexity and downstream accuracy, with a decode kernel that matches or beats FlashAttention; sources include the paper, GitHub repo, shortlink, and a duplicate arXiv reference.
• Ted Zadouri presented FlashAttention-4 at GPU MODE, a Blackwell-focused attention algorithm and kernel that tackles shifted bottlenecks in softmax, memory movement, shared-memory bandwidth, and long-sequence workloads; the batch also included a YouTube URL with tracking parameters and a related X post.
• Romain Lopez and team released icCITE-plex, DOGMA-plex, MoCAVI, and PERCISTRA for massively multiplexed chemical screens at single-cell resolution, profiling RNA, 477 surface proteins, 64 intracellular epitopes, and chromatin accessibility across about 410K primary T cells and 2,800 compound-dose conditions; the thread highlighted mechanisms, off-target effects, and TF-driven chromatin-to-transcription links.
• Rohan Paul summarized Yann LeCun's LeJEPA paper, saying it learns hidden world variables reliably only when those causes follow a balanced Gaussian structure; the batch included a shortlink to the paper.
• Scaling Laws for Agent Harnesses, with social discussion links from omarsar0 and ariG23498, argued that agent performance depends less on raw tokens, tool calls, wall time, or cost and more on “effective feedback compute,” meaning whether the feedback an agent receives is informative, valid, non-redundant, and retained for later decisions; in the paper, raw tokens and tool calls explained limited variation (R² 0.33 and 0.42), while the strongest EFC/task-demand version reached R² 0.99 and fixed-budget feedback improvements raised success from 0.27 to 0.90.
• Reflective Prompt Tuning showed function-calling LLMs acting as prompt engineers: diagnosing failure modes, clustering patterns, generating reports, revising prompts with memory, and improving multi-hop QA and math reasoning; Farima Fatahi shared the thread.
• Learn from your own latents and not from tokens, with related social links from Louis Kirsch and Alessandro Favaro, argued that models can learn more efficiently by predicting their own internal representations of related views or masked regions instead of only predicting visible tokens; the paper proves token-level learning can need samples exponential in the hidden-tree depth L, while latent prediction can recover the same structure with samples constant in L up to logarithmic factors, and suggests explicitly stacking H-JEPA-style hierarchies may be largely redundant. Yichuan Wang paired the research mood with a business take: RL-as-a-Service is probably weak right now because enterprise ROI still favors prompt engineering, RAG, memory, tool use, and agent harnesses over slow, operationally heavy post-training loops.
• OmniRetrieval proposed a unified retrieval layer for mixed knowledge sources, with social context from gabriell_lab, Tomasz Tunguz, Keshigeyan, Jinheon Baek, Guinness Chen, and Aparna Dhinakaran; the paper's core idea is that real-world answers live across text, tables, relational databases, knowledge graphs, and property graphs, so the system routes a natural-language question to the right source and native query engine instead of flattening everything into one retrieval method, beating single-source baselines across 13 datasets and 309 knowledge bases.
• FML-bench studied AI research-agent strategies through search dynamics, with Zhengyao Jiang linking the discussion.
• Sakana AI Labs introduced DiffusionBlocks, an ICLR 2026 method for training neural networks block by block through a diffusion interpretation, reducing memory needs while matching end-to-end performance across ViT, DiT, and transformers.
• LEGO Diffusion from UT Austin and ByteDance uses stackable and skippable blocks for variable-resolution diffusion modeling, combining patch-wise local MLP processing, global Transformer processing, adaptive skipping during sampling, and progressive training.
• HuggingPapers highlighted NVIDIA's quantized Qwen3.6 MoE model, 35B total parameters and 3B active, using NVFP4 quantization to shrink memory by about 3x with near-zero accuracy loss and multimodal support to 262K context; Harvey and Trajectory Labs also post-trained NVIDIA's open-weight Nemotron 3 Super on a Legal Agent Benchmark covering 1,200+ complex legal tasks across 24 practice areas, reportedly matching or approaching closed-source frontier models while preserving auditability and data sovereignty.
• CBS News covered “Dreams of Violets,” an AI-generated film that reportedly cost $2,000 and took two months to make with no lights, cameras, or actors.

🏛️ AI Policy, Governance & Safety

• Foreign Affairs argued China is running an “AI heist” through unauthorized distillation of U.S. frontier models, creating cheap open-weight local models and forcing the U.S. to combine targeted controls with faster open-weight progress.
• CODA, representing major Japanese IP holders, warned tech companies that using anime and manga for AI training is a serious rights issue under Japanese law because models can produce identical or strikingly similar outputs.
• The Washington Post opinion page argued that U.S. AI supremacy depends on both pioneering technology and preventing Chinese AI from conquering global markets.
• CNN warned that AI voice-cloning scams are rising, citing a California mother who lost thousands after receiving a call that sounded like her daughter in distress.
• America Magazine argued that students booing AI-praising commencement speeches is hopeful only if young people also make the smaller daily choice to do their own work instead of outsourcing it to AI.
• FT reported UK chief secretary to the Treasury Lucy Rigby warning that avoiding AI in public services would mean “choosing decline.”
• Gray Swan, an AI security company trusted by major frontier labs, raised $40M Series A to scale enterprise AI security and safe-use products.

🛠️ AI Tools & Products

• Ideabrowser surfaces daily startup ideas, vetted idea databases, trending keywords, market reports, and go-to-market tactics so founders can spot opportunities faster.
• Sinalytica is Sina Rajaeeian's retro demo for running Lovable inside a Windows 98-style environment, with the portfolio site showing his broader AI / ML, web, mobile, and quant-finance work; the batch also included a generic X home link.
• Open Design demonstrated Claude Opus 4.8 recreating a complex Three.js hover-card experiment with interaction, lighting, materials, and cube geometry intact, inspired by an original Cursor build; sources included a shortlink and the Open Design GitHub repo.
• Michael Rabinovich tested Claude Opus 4.8 on real CAD modeling tasks with build123d and showed mixed results versus 4.6 and 4.7 on precise 3D geometry.
• Kai resurfaced “Putting Bugs in Your DC Might Actually Be a Good Idea,” a 2020 paper proposing spiders and silk webs as a self-sustaining way to improve data-center airflow, with the full PDF included.
• Mediagazer remains a single-page aggregator for must-read media news and useful monitoring context.
• Fast Company listed 20 practical Gemini uses, from remembering random facts to writing spreadsheet formulas and decoding product manuals.
• Notion, novasarc01, testingham, Google, and Pope Francis contributed lower-context social items in the batch, preserved here because they informed the weekend's wider product and public-reaction picture.

📊 Fundraising & Deals Roundup

• Groq is reportedly raising $650M in internal funding as it pivots toward inference after Nvidia's failed $20B not-acqui-hire.
• XCENA raised $135M at a $570M valuation to attack AI's memory bottleneck.
• Inherent, an ex-DeepMind AI science startup, raised $50M backed by Index, with Matt Clifford advising; the batch also included the company's site, which frames the lab around recursively self-improving discovery.
• Gray Swan raised $40M Series A for enterprise AI security used by frontier labs.
• Polsia claimed a $30M raise at a $250M valuation while nearing $10M ARR, framed as one founder plus AI and zero employees, with a shortlink to the company site.
• Triomics raised $22M Series B led by Battery Ventures to bring oncology-specific AI into cancer centers.
• General Compute raised $15M seed to build an inference-focused neocloud around SambaNova chips.
• Minute Media laid off 12% of its workforce and reversed its VideoVerse integration after a $200M acquisition, citing efficiency needs and AI-driven disruption in media.

🎙️ Interviews, Panels & Podcasts

• Gemini co-leads, including Jeff Dean and Koray Kavukcuoglu, sat down with Logan Kilpatrick to discuss Gemini 3.5 Flash's origins and what comes next.
• Palantir's AIP Evolve demo showed Chad and Colton using agents to autonomously swap models, tune prompts, and reduce agent costs.
• FlashAttention-4 got a full GPU MODE presentation from Ted Zadouri, useful for readers tracking the low-level systems work behind faster model serving.

💡 Industry Commentary & Analysis

• Caleb Gross argued in “You can just say it” that the AI-era debate keeps moving the goalposts for why humans are valuable, when the cleaner claim is that humans have inherent value regardless of what AI can replicate; the original X post carried the shorter version.
• Bill Gurley argued you do not need “nefarious accounts” to explain AI backlash when the founder of a leading model company just met the Pope and said many of the same negative things; The Washington Post's social post carried the related data-center claim story.
• Lisan al Gaib predicted OpenAI and Anthropic will stay at the frontier, Google and Chinese labs will struggle to catch up on long-horizon coding and R&D, top intelligence will become a $999+ luxury good, AI investors will win, SpaceX AI will match Google by year-end, and Nvidia will reach $10T first.
• Gary Marcus offered the opposing hot take: tokenmaxxing will decline, OpenAI and Anthropic will struggle with profit, Google and some Chinese labs may catch up, LLMs will commoditize, big investors may lose, SpaceX AI will flail, and Nvidia will eventually decline.
• Epoch AI Research posted a gap analysis on open versus closed model progress.
• llama.app became the official home for llama.cpp, and Georgi Gerganov said it now offers a single-line cross-platform installer, a unified llama entrypoint for running and serving models, agentic app interfaces, and automatic reuse of existing GGUF models in the Hugging Face cache.
• SemiAnalysis noted that many of AMD and NVIDIA's strongest 10x engineers are in Shanghai, highlighting the city's importance in AI hardware talent.
• Harry McCracken argued AI-assisted journalism needs explicit disclosure, sharing his own rule of never pasting chatbot text directly into drafts and always tracing suggested quotes back to original sources.
• Eve Fairbanks argued the biggest tell of AI writing is that every part of the text is “not quite right,” because it can be grammatical, bland, oddly structured, and hard to edit without an underlying human reasoning process.
• TechCrunch argued Paris may now be the most important AI city outside Silicon Valley because European founders increasingly scale domestically instead of immediately moving to the U.S.
• Mistral's CEO warned that Europe's main obstacle to AI independence is the investment scale needed to chase superintelligence and counter U.S. dominance.
• Al Jazeera reported that China's cheap electricity is a major advantage in the AI data-center buildout.

That's a Wrap

That's 150+ weekend links across models, agents, chips, policy, jobs, and the strange new science of making agents stop spending money like a corporate card with no receipt policy. If you made it this far, congratulations, you now have enough context to nod thoughtfully the next time someone says “agentic” in a meeting and then invoices $999/month for the privilege.

For the daily version, make sure you're subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you do not have to.

See you tomorrow.

P.S: Know someone who'd find this useful? Forward this to them and tell them to subscribe here.