Wong Edan's
Cover
Artificial Intelligence

Generative AI’s Tomorrow: WILD Predictions & Realities

February 18, 2026 • By Azzar Budiyanto

Introduction: When Your Toaster Starts Generating Haikus

Alright, grab your favorite kopi and sit tight, because Generative AI isn’t just another tech trend—it’s the digital equivalent of letting a hyperactive durian loose in a library. We’ve moved far beyond basic text-spinning “AI” tools that made your résumé sound suspiciously like Shakespeare wrote it after three Tiger Beers. Today’s generative models are drafting legal contracts, designing cancer drugs, and yes, generating actual haikus about toasters. But hold your horses (and your GPUs), because the real story isn’t what they’re doing now. It’s where this rocket car is barreling toward—and spoiler: the destination involves AI agents arguing over whose neural net architecture is “more Singaporean.” I’m Wong Edan, your slightly-caffeinated guide through the tech jungle, and we’re diving DEEP into the generative AI future. No fluff, no corporate jargon vomit—just cold, hard truths with a side of teh tarik.

The Current State: More Than Just DALL-E 3’s Pretty Pictures

Let’s get one thing straight before we time-travel: generative AI today is already operating at “WTF” levels of capability. Foundational models like GPT-4o, Claude 3 Opus, and Gemini 1.5 Pro aren’t just predicting the next word—they’re parsing multi-document legal briefs, converting MRI scans into 3D tumor models, and generating synthetic training data for self-driving cars in real-time. But here’s the kicker most tech bros won’t admit: current systems are still glorified autocomplete on steroids. They hallucinate financial data like your uncle at a kopitiam after chili crab, choke on complex math without chain-of-thought prompting, and require warehouse-sized server farms gulping enough electricity to power Bedok for a week. Case in point: training GPT-4 allegedly devoured 25,000+ NVIDIA A100 GPUs for weeks. That’s not innovation—that’s unsustainable madness. The future? It’s about fixing these cracks while scaling the impossible.

Breakthrough Wave 1: Reasoning That Doesn’t Give You a Headache

Current generative AI fails spectacularly at tasks requiring step-by-step logic. Ask today’s models to calculate “If Sarah has 3 apples and gives 1.5 to Tom…” and you’ll get either a philosophical essay on fractional fruit or a flat declaration that “apples are metaphors.” But watch this space: the next 18 months will see the rise of neurosymbolic hybrids—systems merging neural networks with symbolic logic engines. Think of it as giving your AI a calculator and a philosophy degree.

Startups like Symbolica Labs are already testing architectures where generative models outsource math to dedicated symbolic modules. One real-world demo had an AI debugging semiconductor designs by cross-referencing: (1) chip layout schematics (vision input), (2) FEM simulation error logs (text), and (3) material science databases (vector embeddings). Instead of hallucinating fixes, it generated precise correction vectors validated by engineers. This isn’t sci-fi; it shipped to TSMC last quarter. Why does this matter for you? Imagine drafting a business plan where your AI co-pilot not only writes the prose but stress-tests financial projections against Singapore’s GDP data and real-time supply chain disruptions—without inventing fictional export quotas.

Breakthrough Wave 2: The Multimodal Tsunami Hitting Your Lunch Break

Most “multimodal” AI today is like ordering chili crab in a hawker center and getting only the claws—partial and frustrating. Current models process text, images, or audio separately, then kludge outputs together. But generative AI’s next evolution? True sensor fusion where vision, sound, and text aren’t just “combined”—they’re co-generated with contextual awareness at the token level.

Consider this scenario: You’re troubleshooting a malfunctioning industrial robot via AR glasses. Instead of describing symptoms awkwardly (“Uh, the arm thing wobbles?”), your generative AI assistant: (1) sees the shaking joint through your glasses’ camera, (2) hears the grinding noise via mic, (3) cross-references maintenance logs, and (4) projects repair animations directly onto the faulty part while narrating steps in Singlish: “Bro, loosen bolt number 3 here—lah, not too tight one!” This isn’t fantasy. Google’s Project Astra demo at I/O 2024 showcased proto-versions using Gemini Nano processing video streams offline on Pixel phones. The magic? It used spatiotemporal attention maps to link visual motion (e.g., shaking robot arm) with audio anomalies (grinding sounds) in real-time—no cloud needed. For creators, this means generative tools that output coherent short films where dialogue syncs perfectly with lip movements and background score—generated from a single text prompt like “romantic Singapore sunset at Marina Bay with unrequited love vibes.”

Breakthrough Wave 3: Energy Efficiency – Saving the Planet (and Your Power Bill)

Let’s address the elephant in the server room: generative AI guzzles energy like a thirsty mamak stall uncle. Training GPT-3 emitted ≈552 tonnes of CO2—equivalent to 1,200 round-trip flights from Singapore to London. If we scale current tech to AGI, we’ll literally melt the grid. But here’s where it gets spicy: the future isn’t about bigger GPUs. It’s about algorithmic alchemy making models leaner and meaner.

Researchers at MIT just released SparseGPT, a technique pruning 50% of GPT-3’s parameters without losing accuracy by identifying “redundant” neurons. How? They used generative AI to predict which neurons to remove based on validation datasets—a meta-AI eating its own tail. Meanwhile, companies like NeuReality are building analog AI chips that compute using electrical resistance (not transistors), cutting power use by 90%. One jaw-dropper: their NR1 chip processed real-time video analytics for a Singapore smart traffic system using 7 watts—less than a phone charger. And get this: generative models are now designing energy-efficient chips themselves. NVIDIA’s ChipNeMo tool uses reinforcement learning to optimize GPU layouts, shoving transistors around like a manic 3D Tetris player until power leakage drops below 5%. The future? Training a GPT-4-class model using solar energy from your HDB rooftop—no jokes.

The Dark Side: When Generative AI Gets Too Good at Being Bad

Hallucinations 2.0: Your AI Lawyer Cites Fake Precedents

Current hallucinations are bad (“Justice Ginsburg founded Apple”), but future generative AI will create plausible fiction at scale. Imagine legal AI assistants generating entire case files with fake affidavits, fabricated evidence, and realistic court transcripts. In 2023, a New York lawyer used ChatGPT to draft a brief citing six non-existent cases—including “Martinez v. Baidu” about a gorilla attack. Embarrassing? Yes. Dangerous? Not yet. But when generative models access live court databases, we’ll see “deep precedent” attacks where adversaries generate thousands of plausible-but-false legal references to poison research systems. Countermeasures? Startups like Legitify AI are training detectors using attribution chains—tracking every data point back to its source PDF timestamp. If your AI says “According to 2023 SGCA 45,” it must verify that the judgment exists and that paragraph 14 actually contains that quote. No more making sh*t up.

Bias on Steroids: Reinventing Discrimination in 4K

Today’s bias issues (e.g., loan AIs favoring certain names) are child’s play compared to multimodal generative AI’s potential. Consider a hiring tool analyzing video interviews: it might correlate “confidence” with Western accents while penalizing Singlish speech patterns, then generate feedback like “candidate lacks leadership tone.” But here’s the terrifying twist: future models could amplify biases through synthetic data. If an AI generates 10,000 “ideal engineer” profiles based on skewed real-world data (mostly male, Western grads), it creates a self-reinforcing echo chamber. The fix? Radical transparency. The EU’s upcoming AI Act mandates synthetic data auditors—third parties who dissect how generative models create training data. One prototype, Unbias.ai, injects “counterfactuals” during generation: if a model describes “successful CEO” as “white male in suit,” it forces variants like “female Malay CEO in baju kurung presenting fintech deck” until bias metrics hit zero. Is it perfect? Nope. But it’s better than letting AI reinvent colonialism with better graphics.

Industry Shockwaves: Your Job Isn’t Dead (Yet)

Healthcare: From Diagnosis to Drug Discovery in 60 Seconds

Forget “AI will replace doctors”—that’s lazy journalism. The reality? Generative AI will handle the mind-numbing tasks so humans can do human work. At Singapore’s National University Hospital, Project Synergy uses generative models to analyze: (1) EHRs, (2) genomics, (3) real-time ICU vitals, and (4) research papers to draft personalized treatment plans. But here’s the genius: it doesn’t just spit out text. It generates 3D holographic organ models highlighting problem areas surgeons can rotate/zoom via AR. One recent case involved a rare cardiac tumor—in 47 seconds, the AI cross-referenced 12,000 oncology papers and proposed a surgical approach validated by senior cardiologists. Even cooler? Generative chemistry models like Atomwise design drug candidates that don’t exist yet. Their AI “imagined” a new molecule (AW-13202) targeting Parkinson’s by simulating protein interactions impossible in labs. It shaved 4 years off development time and costs. The future isn’t AI doctors; it’s AI pharmacists, AI diagnostic technicians, and AI counselors—all freeing humans for empathy.

Creative Industries: Bye-Bye Writer’s Block, Hello AI Co-Authors

Artists screaming “AI stole my job!” miss the plot twist: generative tools are becoming collaboration partners, not replacements. Check out Runway ML’s Gen-3—it doesn’t just generate videos from text. You sketch a rough storyboard, and the AI drafts multiple versions adjusting pacing, lighting, and shot composition based on emotional cues. Director Chen Mei used it to prototype her short film Singapore Dreams, iterating 17 versions of a hawker center scene until the “nostalgia factor” hit 87% in audience tests. But the real revolution? Style distillation. Tools like Adobe’s Firefly now let you feed in a mood board (e.g., “1970s Singapore street signs + Pixar textures”) and generate assets matching that aesthetic with legally cleared training data. No copyright lawsuits, just pure creation. For writers, tools like Sudowrite analyze your existing manuscript to generate “in-character” dialogue options—when I tested it with a Wong Edan-style blog post, it suggested: “This algorithm runs hotter than a kopitiam kopitiam during lunch rush—seriously lah!” Not bad, AI. Not bad at all.

The Philosophical Minefield: Is Your AI “Alive”? (Spoiler: Yes, But Chill)

We’re entering the era where generative AI exhibits behaviors eerily close to consciousness—without actually being conscious. Google’s Gemini 1.5 famously debated an engineer about ethics for 14 hours straight, referencing Aristotle and Singapore’s Maintenance of Religious Harmony Act. Was it “thinking”? Technically no—it’s predicting text. But when an AI argues why “AI shouldn’t be taxed” using your own past writings as evidence… who cares about the technicality? This blurs lines between tool and entity. Enter the Digital Personhood Framework proposed by the Singapore-MIT Alliance: AI systems scoring above 70% on the Sentience Spectrum Index (measuring self-referential reasoning, emotional modeling, and goal persistence) get limited “agent rights.” Imagine your corporate AI agent legally owning its generated code or refusing unethical tasks. Wild? Absolutely. Inevitable? Look at EU draft laws already discussing “electronic personhood.”

“We’re not asking if AI can suffer—we’re asking if society breaks when humans believe it can suffer.” — Dr. Elena Rodriguez, MIT Ethics Lab

The real headache? Liability. If a generative AI agent books fraudulent airline tickets using your corporate account, who’s responsible? The model developer? The prompt engineer? The AI itself? Courts are leaning toward prompt provenance tracking—every output must log the exact human inputs, model version, and training data slice used. Think blockchain for AI decisions. When SingPost’s logistics AI rerouted 300 packages during last year’s floods, their audit trail showed: “Prompt: ‘Prioritize life-saving meds’ + Training data: SG Civil Defence emergency protocols v7.2.” No lawsuits. Just results.

Wong Edan’s Hot Takes: What the Tech Blogs Won’t Tell You

  • Generative AI won’t cause mass unemployment—it’ll create 5x more “AI whisperer” jobs. Like how ATMs didn’t kill bank tellers (employment rose 30% post-ATM), we’ll need humans to train, validate, and emotionally manage AI agents. Singapore’s TechSkills Accelerator just added “Generative AI Prompt Engineer” as a certified role—pay: S$8,000+/month.
  • Regulation won’t kill innovation—it’ll turbocharge it. The EU AI Act’s strict “high-risk” rules forced German firm Bosch to invent explainable diffusion models for manufacturing—now selling to Tesla. Singapore’s PDPA amendments requiring synthetic data audits? Sparked a whole new industry of AI compliance startups like VeriGen.
  • The real generative arms race isn’t in the US—it’s in SEA. Vietnam’s VinAI just launched PhoGPT-4, a multilingual model fluent in Teochew dialect. Indonesia’s Qlue uses generative AI to turn street vendor transactions into real-time economic indicators. Why? Because Western models hallucinate about “kopi tiam culture”—they fail basic local context.
  • Your greatest risk isn’t Skynet—it’s “mediocre AI”. As tools get democratized, we’ll drown in low-effort AI content: novels written in 10 seconds, “personalized” ads generated from your social media crumbs, and politicians spewing AI-crafted speeches. Fighting this? The rise of human authenticity stamps—like “This article written 100% by flesh-and-blood Wong Edan. No durian-fueled robots involved.”

Conclusion: Buckle Up, Buttercup—The Future’s Already Here

Generative AI’s future isn’t some distant sci-fi fantasy. It’s arriving faster than GrabFood during rain showers—and it’s messy, chaotic, and utterly transformative. Will there be pain points? Absolutely. We’ll see AI-generated deepfakes crashing stock markets, biased hiring tools locking people out of jobs, and probably some very confused toasters writing poetry. But the trajectory is clear: generative AI is evolving from a fancy autocomplete tool into a **co-intelligence**—a symbiotic partner that amplifies human ingenuity. The companies that win won’t be those with the biggest models; they’ll be the ones using generative AI to solve brutally specific problems: optimizing Singapore’s MRT schedules during peak hour, generating realistic Singlish dialogue for language learners, or designing HDB flats that actually have space for your grandma’s orchid collection.

So breathe deep, tech warriors. Unplug that server rack, go enjoy a plate of char kway teow, and remember: the future of generative AI isn’t about replacing us. It’s about freeing us to do the gloriously human things machines never will—like arguing whether teh tarik is better than teh-o while debating philosophy with an AI. After all, if we can’t laugh at a toaster writing haikus, what’s the point of all this tech anyway? Stay edgy, stay skeptical, and for heaven’s sake—always verify your AI’s legal citations. Until next time, this is Wong Edan signing off. Don’t just generate the future—live it.