Featured by dilanshah

How can mixed reality drive more engagement in movement and fitness?

Fitness is one of the most robust categories under discussion, across Augmented Reality and Virtual Reality devices. For whom does this movement level merit the moniker “fitness”? And what timeline are we working with to see the sweeping adoption of fitness via spatial computing (the term widely known now due to Apple’s Vision Pro announcement and curving of the terms VR / AR / MR collectively)?

I’m seeing new unlocks particularly as it relates to the comfort of the device, spatial awareness afforded due to camera passthrough, and greater respect for ergonomic polish among developers.

The video seen here is a clip taken November 8th, 2023, showing a first-person view of a Quest 3 experience that allows for gestures, hand tracking, and movement to be used as input to an increasing number of games.

The title is built by YUR, the app name is YUR World.
by dilanshah

What Building Enterprise AI Actually Looks Like Right Now

The hype around generative AI makes it sound like companies simply “turn on” AI and watch the magic happen. In reality, deploying AI inside large organizations is messy, technical, and deeply human.

Through conversations with builders working on enterprise AI deployments, a clearer picture emerges of what the work actually looks like on the ground.

The Marketing vs. Product Reality Gap

AI companies are still figuring out how to talk about what they build. Marketing teams often experiment with language to position their platforms at conferences and events, while product teams sometimes cringe at the terminology being used.

Part of the challenge is that the industry itself is evolving in real time. Everyone—from engineers to executives—is still forming opinions about how AI systems should be described, positioned, and sold. The result is a constant negotiation between what sounds compelling on stage and what accurately reflects the underlying technology.

The Rise of the “Forward-Deployed” Builder

A growing role in enterprise AI is the forward-deployed product or engineering role. These builders sit somewhere between product manager, consultant, and engineer.

Instead of building software in isolation, they embed themselves in real customer deployments. Their job is to translate a customer’s operational reality into something the platform can support.

Typical responsibilities include:

  • Understanding a customer’s existing workflows and systems
  • Designing how AI agents integrate into those environments
  • Testing and validating that systems behave correctly in production scenarios
  • Acting as the bridge between engineering teams and enterprise stakeholders

Each large deployment often involves a small pod—engineers and product specialists—working closely with a single customer.

Enterprise AI Is Still Mostly Strategy

One surprising reality: many large organizations want AI but don’t yet know what to do with it.

Companies frequently arrive with enthusiasm but little internal expertise. In many cases, a single internal champion pushes the initiative forward without a broader AI team behind them.

As a result, AI vendors often end up helping customers define their strategy before they even implement anything. Conversations quickly shift from “Can we deploy AI?” to questions like:

  • Where in our workflows does automation actually make sense?
  • What customer interactions should remain human?
  • How do we measure success for an AI agent?

For many enterprises, AI adoption begins as a strategic exercise rather than a purely technical one.

The Hidden Work: Testing, QA, and Reliability

Much of the real work in AI deployments isn’t glamorous.

Ensuring that AI systems behave reliably—especially in customer-facing contexts—requires extensive testing and quality assurance. Teams spend significant time verifying that agents respond correctly across thousands of scenarios.

For example, a customer-service AI agent might need to correctly handle thousands of variations of a simple request like “Where is my order?” — across different policies, authentication states, and edge cases. Each of those paths needs to be validated before automation can be trusted.

In practice, this means:

  • validating conversation flows
  • testing edge cases and failure modes
  • ensuring integrations work across multiple internal systems
  • implementing human-in-the-loop review processes before deployment

The goal is often to handle the “happy path” of common interactions reliably before expanding to more complex scenarios.

Platform vs. Services: A Tension in the Industry

Another emerging tension in AI companies is the balance between building a scalable platform and delivering hands-on services to customers.

Enterprise AI deployments frequently require significant customization and implementation support. Investors and product leaders, however, typically prefer companies that scale through product rather than services.

Many companies are navigating this hybrid model:
a product platform at the core, with a services layer that helps customers adopt it.

Over time, the hope is that repeated deployments lead to reusable templates and more standardized implementations.

The Next Wave: Personalized Software

One theme that keeps emerging across the AI ecosystem is the idea that we’re entering an era of personalized software.

Instead of buying rigid tools, organizations may soon generate or customize applications tailored to their exact workflows. Early experiments already show teams rapidly building demo environments and prototypes for customers using generative development tools.

If this trend continues, the biggest shifts in the next few years may not be just smarter models—but entirely new ways software gets built and delivered.

The Takeaway

The current wave of enterprise AI isn’t just about models or algorithms. It’s about integrating intelligent systems into messy real-world environments.

That requires a new kind of builder—part product thinker, part engineer, part strategist—working directly alongside customers to turn AI from a concept into something that actually works.

The hype cycle is loud.
But the real work of enterprise AI is happening quietly — inside deployment pods, QA spreadsheets, and integration docs.

by dilanshah

Understanding RNA Therapeutics: Revolutionizing Medicine

RNA therapeutics are medicines that use ribonucleic acid (RNA) molecules as the active drug to treat or prevent disease. Instead of traditional small-molecule drugs or protein-based biologics, they work by directly influencing how genes are expressed inside cells.

There are several main types of RNA therapeutics:

TypeHow It WorksKey Examples (Approved or Famous)Main Uses
mRNA vaccines / therapeuticsDeliver messenger RNA (mRNA) that instructs cells to produce a specific protein (e.g., viral spike protein or a missing enzyme)Pfizer-BioNTech & Moderna COVID-19 vaccines, BioNTech’s cancer vaccines (in trials)Vaccines, cancer immunotherapy, protein replacement
ASOs (Antisense Oligonucleotides)Short, synthetic single-stranded DNA/RNA-like molecules that bind to target mRNA and block or degrade itNusinersen (Spinraza) for spinal muscular atrophy, Inotersen & Patisiran for hereditary ATTR amyloidosisRare genetic diseases, neurological disorders
siRNA (small interfering RNA)Double-stranded RNA that triggers the cell’s natural RNA interference (RNAi) machinery to silence specific genesPatisiran (Onpattro) – first ever FDA-approved siRNA, Givosiran for acute hepatic porphyriaGenetic diseases, liver diseases, some cancers
saRNA (self-amplifying RNA)mRNA that encodes not only the target protein but also a viral replicase, so it copies itself inside the cell → longer, stronger protein production with tiny dosesIn development (e.g., Gritstone, Arcturus COVID/flu programs)Vaccines (especially low-dose, fridge-stable ones)
RNA aptamersFolded RNA molecules that bind proteins like antibodiesPegaptanib (Macugen) – first RNA aptamer drug (for macular degeneration)Eye diseases, anticoagulation, cancer
Circular RNA (circRNA)RNA in a closed loop → very stable, long-lasting protein expressionEarly clinical trials (e.g., Orna Therapeutics)Protein replacement, vaccines

Why RNA Therapeutics Are a Big Deal

  1. Speed of development
    – COVID mRNA vaccines went from sequence to emergency use in <1 year (vs. 10–15 years for traditional vaccines).
  2. Precision
    – You can target almost any gene or protein. If we know the genetic cause of a disease, we can design an RNA drug against it.
  3. “Undruggable” targets become druggable
    – Many diseases are caused by proteins that small molecules can’t bind well. RNA drugs act before the protein is even made.
  4. Personalization potential
    – Easy to customize mRNA sequence for a patient’s specific mutation (already happening in cancer vaccines).

Major Challenges (Why They’re Hard)

ChallengeExplanationCurrent Solutions / Progress
DeliveryNaked RNA is destroyed quickly by enzymes and can’t easily enter cellsLipid nanoparticles (LNPs), GalNAc conjugates, new polymers
Immune activationRNA can trigger strong inflammatory responsesChemical modifications (pseudouridine, etc.)
Manufacturing scale-upVery sensitive biologic; hard to make consistently at huge scaleMassive investment post-COVID; new platforms emerging
Duration of effectMost RNA effects are transient (days to weeks)circRNA, saRNA, repeated dosing, or gene editing combos
CostStill expensive compared to small-molecule pillsEconomies of scale improving rapidly

The Future (2025–2030 Outlook)

  • Hundreds of RNA programs in clinical trials (cancer, rare diseases, infectious diseases, Alzheimer’s, heart disease, etc.).
  • Next-generation delivery: targeting lungs, brain, heart muscle, tumors directly.
  • Combination with CRISPR: using mRNA to deliver gene-editing machinery (already in trials).
  • Off-the-shelf and personalized cancer vaccines likely to get approved in the next few years.

In short: RNA therapeutics are one of the fastest-growing areas in medicine right now. They turned science fiction (programmable medicines) into reality with the COVID vaccines, and the pipeline behind them is enormous.

by dilanshah

Elon Musk, Diffusion Models, and the Rise of Mercury

  1. A New Paradigm May Be Forming
  2. Meet Inception Labs and Mercury
  3. How Mercury Works
  4. Inside the Diffusion Revolution
  5. Training and Scale
  6. Performance: 10× Faster, Same Quality
  7. A Historical Echo
  8. What Comes Next
  9. Further Reading

A New Paradigm May Be Forming

In a recent exchange on X, Elon Musk echoed a striking prediction: diffusion models — the same architecture that powers image generators like Stable Diffusion — could soon dominate most AI workloads. Musk cited Stanford professor Stefano Ermon, whose research argues that diffusion models’ inherent parallelism gives them a decisive advantage over the sequential, autoregressive transformers that currently power GPT-4, Claude, and Gemini.

While transformers have defined the past five years of AI, Musk’s comment hints at an impending architectural shift — one reminiscent of the deep learning revolutions that came before it.

Meet Inception Labs and Mercury

That shift is being engineered by Inception Labs, a startup founded by Stanford professors including Ermon himself. Their flagship system, Mercury, is the world’s first diffusion-based large language model (dLLM) designed for commercial-scale text generation.

The company recently raised $50 million to scale this approach, claiming Mercury achieves up to 10× faster inference than comparable transformer models by eliminating sequential bottlenecks. The vision: make diffusion not just for pixels, but for language, video, and world modeling.

How Mercury Works

Traditional LLMs — whether GPT-4 or Claude — predict the next token one at a time, in sequence. Mercury instead starts with noise and refines it toward coherent text in parallel, using a denoising process adapted from image diffusion.

This process unfolds in two stages:

  1. Forward Process: Mercury gradually corrupts real text into noise across multiple steps, learning the statistical structure of language.
  2. Reverse Process: During inference, it starts from noise and iteratively denoises, producing complete sequences — multiple tokens at once.

By replacing next-token prediction with a diffusion denoising objective, Mercury gains parallelism, error correction, and remarkable speed. Despite this radical shift, it retains transformer backbones for compatibility with existing training and inference pipelines (SFT, RLHF, DPO, etc.).

Inside the Diffusion Revolution

Mercury’s text diffusion process operates on discrete token sequences x \in X. Each diffusion step samples and refines latent variables z_t that move from pure noise toward meaningful text representations. The training objective minimizes a weighted denoising loss:

L(x) = -\mathbb{E}t [\gamma(t) \cdot \mathbb{E}{z_t \sim q} \log p_\theta(x | z_t)]

In practice, this means Mercury can correct itself mid-generation — something autoregressive transformers fundamentally struggle with. The result is a coarse-to-fine decoding loop that predicts multiple tokens simultaneously, improving both efficiency and coherence.

Training and Scale

Mercury is trained on trillions of tokens spanning web, code, and curated synthetic data. The models range from compact “Mini” and “Small” versions up to large generalist systems with context windows up to 128K tokens. Inference typically completes in 10–50 denoising steps — orders of magnitude faster than sequential generation.

Training runs on NVIDIA H100 clusters using standard LLM toolchains, with alignment handled via instruction tuning and preference optimization.

Performance: 10× Faster, Same Quality

On paper, Mercury’s numbers are eye-catching:

BenchmarkMercury Coder MiniMercury Coder SmallGPT-4o MiniClaude 3.5 Haiku
HumanEval (%)88.090.0~8590+
MBPP (%)76.677.1~75~78
Tokens/sec (H100)110973759~100
Latency (ms, Copilot Arena)25N/A~100~50

Mercury rivals or surpasses transformer baselines on code and reasoning tasks, while generating 5–20× faster on equivalent hardware. Its performance on Fill-in-the-Middle (FIM) benchmarks also suggests diffusion’s potential for robust, parallel context editing — a key advantage for agents, copilots, and IDE integrations.

A Historical Echo

Machine learning has cycled through dominant architectures roughly every decade:

  • 2000s: Convolutional Neural Networks (CNNs)
  • 2010s: Recurrent Neural Networks (RNNs)
  • 2020s: Transformers

Each leap offered not just better accuracy, but better compute scaling. Diffusion may be the next inflection point — especially as GPUs, TPUs, and NPUs evolve for parallel workloads.

Skeptics, however, note that language generation’s discrete structure may resist full diffusion dominance. Transformers enjoy massive tooling, dataset, and framework support. Replacing them wholesale won’t happen overnight. But if diffusion proves cheaper, faster, and scalable, its trajectory may mirror the very transformers it now challenges.

What Comes Next

Inception Labs has begun opening Mercury APIs at platform.inceptionlabs.ai, pricing at $0.25 per million input tokens and $1.00 per million output tokens — a clear signal they’re aiming at OpenAI-level production workloads. The Mercury Coder Playground is live for testing, and a generalist chat model is now in closed beta.

If Musk and Ermon are right, diffusion could define the next chapter of AI — one where text, video, and world models share the same generative backbone. And if Mercury’s numbers hold, that chapter may arrive sooner than anyone expects.

Further Reading

  • Stefano Ermon et al., Diffusion Language Models Are Parallel Transformers (Stanford AI Lab)
  • Elon Musk on X, Diffusion Will Likely Dominate Future AI Workloads
  • Inception Labs, Mercury Technical Overview (2025)

by dilanshah

Build and Sell n8n AI Agents — an eight-plus-hour, no-code course

AI Agents and Automation

  1. AI agents have two parts: a brain—that is, a large language model with memory—and instructions in the system prompt. Together they let the agent make decisions and take actions through connected tools.
  2. Reactive prompting beats proactive prompting: begin with no prompt, then add lines only when errors appear. This makes debugging simpler.
  3. Give each user a unique session ID so the agent’s memory stays separate, enabling personal conversations with many users at once.
  4. Use Retrieval-Augmented Generation, or R-A-G. The agent asks a question, looks up an answer in a vector database, then crafts the reply—boosting accuracy.

AI Workflows and Best Practices

  1. AI workflows—straight, deterministic pipelines—are usually cheaper and more reliable than free-roaming agents, and they’re easier to debug.
  2. Wire-frame the whole workflow first. Mapping eighty to eighty-five percent of the flow upfront clarifies what to build.
  3. Combine agents in a multi-agent system: an orchestrator assigns tasks to specialist sub-agents. That raises accuracy and control.
  4. Apply an evaluator–optimizer loop. One component scores the output; another revises it, repeating until quality is high.

AI Integration and Tools

  1. n8n is a powerful no-code platform for AI automations; you can create and even sell more than fifteen working examples.
  2. Open Router picks the best large language model for each request on the fly, balancing cost and performance.
  3. Eleven Labs adds voice input to an email agent. Pair it with Google Sheets for contacts and the Gmail API for sending mail.
  4. Tavly offers a thousand free web searches per month—handy for research inside AI content workflows.

AI Agent Development Strategies

  1. Scale vertically first: perfect one domain—its knowledge base, data sources, and monitoring—before branching out.
  2. Test rigorously, add guard-rails, and monitor performance continuously before you hit production.
  3. Use hard prompting: spell out examples of correct and incorrect behavior right in the system prompt.
  4. Allow unlimited revision loops when refining text, so the workflow can keep improving its answer until it satisfies you.

AI Business Applications

  1. Three-quarters of small businesses already use AI; eighty-six percent of adopters earn over one million dollars in annual AI-driven revenue.
  2. AI-guided marketing lifts ROI by twenty-two percent, while optimized supply chains trim transport costs five to ten percent.
  3. AI customer-service agents cut response times sixty percent and solve eighty percent of issues unaided.
  4. The median small business spends just eighteen-hundred dollars a year on AI—under one-fifty a month.

AI Development Techniques

  1. Structure prompts with five parts: overview, tools, rules, examples, and closing notes.
  2. Debug one change at a time—alter a single line to isolate the issue.
  3. Log usage and cost in Google Sheets to track tokens and efficiency.
  4. Use polling in workflows: check task status at intervals before moving on.

AI Integration with External Services

  1. In Google Cloud, enable the Drive API, set up OAuth, and link n8n for file workflows.
  2. Do the same with the Gmail API to trigger flows and send replies.
  3. Build a Pinecone vector index (for example, with text-embedding-3-small) for fast R-A-G look-ups.
  4. Generate graphics through OpenAI’s image API to save about twenty minutes per post.

Advanced AI Techniques

  1. Use a routing framework to classify inputs and dispatch them to the right specialist agent.
  2. Add parallelization so different facets of the same input are analyzed simultaneously, then merged.
  3. Store text as vectors in a vector database for semantic search—meaning matters more than keywords.
  4. Deploy an M-C-P server as a universal translator between agents and tools, exposing tool lists and schemas.

AI Development Challenges and Considerations

  1. Remember: most online agent demos are proofs of concept—not drop-in, production-ready templates.
  2. Security matters; an M-C-P server could access sensitive resources, so lock it down.
  3. Weigh agents versus workflows; use agents only when you need complex reasoning and flexible decisions.
  4. Supply high-quality context—otherwise you risk hallucinations, tool misuse, or vague answers.

AI Tools and Platforms

  1. Alstio Cloud manages open-source apps like n8n for you—install, configure, and update.
  2. Tools such as Vellum and L-M Arena let you compare language-model performance head-to-head.
  3. Supabase or Firebase cover user auth and data storage in AI-enabled web apps.
  4. In self-hosted n8n, explore community nodes—for instance, Firecrawl or Airbnb—to expand functionality.
by dilanshah

🚨 The AI Cyber‑Warfare Threat: Insights from Geoffrey Hinton on DOAC

In his appearance on The Diary Of A CEO with Steven Bartlett, Geoffrey Hinton—the so-called “Godfather of AI”—issued a compelling warning about AI’s dual-use potential. While AI offers immense benefits, “at least half” of its development is likely directed towards offensive cyber operations. This includes crafting more potent attacks, designing new malware, and automating exploits in real time.

  1. Cyber‑Attacks Supercharged by AI
    • From reactive to proactive: AI not only defends networks but also enables automated scouting for vulnerabilities and weaponized code generation.
    • Escalating sophistication: Cyber‑criminals and state actors are already leveraging AI to build advanced phishing campaigns. They are also using it to develop malware. This forces a continuous escalation in cyber warfare.
  2. Biological Risks: AI‐Designed Viruses

Hinton raised the specter of AI-aided bioengineering. This crossover risk—where cyber AI knowledge facilitates biological threats—represents a chilling frontier.:

“There’s people using it to make nasty viruses” .

  1. Election Manipulation Beyond Digital Borders
  • AI’s ability to model and influence human behavior isn’t limited to malware. According to Hinton, AI-driven tools can:
  • Craft hyper-personalized messaging to sway individuals,
  • Potentially manipulate public opinion and democratic processes.
  1. Urgent Call for Safety‑First Governance

Hinton emphasized that the moment to act is now:

  • Governments should mandate major AI firms to allocate a portion of compute resources toward safety testing,
  • This includes rigorous safety evaluations prior to release and independent oversight .
  • Without safeguards, profits and power will continue to outweigh safety—leaving us vulnerable.

📝 What a Responsible Defense Looks Like

If you’re thinking about policy and strategic frameworks, here’s a roadmap inspired by Hinton’s analysis:

Key Focus Area Recommended Action

  • Regulation & Oversight Governments must require safety audits for AI models before deployment.
  • Safety‑first R&D Major AI labs should allocate dedicated compute to adversarial safety research.
  • Global Cooperation Collaboration across countries to counter cross-border misuse including bio‑threats.
  • Public Awareness Inform citizens and organizations about AI-driven threat evolution—phishing, malware, targeted political influence.

Final Thoughts

Hinton’s warnings aren’t speculation; they’re grounded in current tech trajectories. AI isn’t just a tool; it’s fast becoming the weapon of choice in cyber and bio conflict.

But there is hope. With proactive safety commitments, regulations tailored to dual-use risk, and global collaboration, we can choose to channel AI’s power responsibly. The question is: will society act before technology outruns us?

by dilanshah

How Direct File is Quietly Redefining Government Software

In a rare but powerful move, the U.S. government has open-sourced one of its most impactful digital public services: Direct File, a platform that allows taxpayers to file their federal returns electronically—completely free of charge and without third-party intermediaries.

At a glance, Direct File might seem like just another government web form. But beneath the surface lies a thoughtfully engineered system that’s not just about taxes—it’s a case study in modern government software, scalable infrastructure, and user-first design.

Let’s break it down.

🧾 What is Direct File?

Direct File is a web-based, interview-style application that guides users through the federal tax filing process. It works seamlessly across devices—mobile, desktop, tablet—and is available in both English and Spanish.

Built to accommodate taxpayers with a wide range of needs, it translates the complexity of IRS tax code into plain-language questions. On the backend, it connects with the IRS’s Modernized e-File (MeF) system via API to handle real-time tax return submissions.

🧠 The Tech Stack: Government Goes Modern

The project reflects a significant leap forward in how federal systems are built and deployed.

  • Fact Graph: At the heart of Direct File is a “Fact Graph”—an XML-based knowledge graph that smartly handles incomplete or evolving user information during the filing process.
  • Programming Stack:
    • Scala for the logic and backend (running on the JVM)
    • Transpiled to JavaScript for client-side execution
    • React frontend in the df-client directory
    • Containerized for Speed: Docker is used for seamless local deployment.
      • This spins up the backend (port 8080) and Postgres DB (port 5432).
  • Modular Architecture:
    • fact-graph-scala: Core tax logic
    • js-factgraph-scala: Frontend port
    • backend: Auth, session management
    • submit: MeF submission engine
    • status: Monitors submission acknowledgments
    • state-api: Bridges federal and state systems
    • email-service: Handles user notifications

🤝 Built by Public Servants, Not Contractors Alone

Unlike many large-scale federal tech initiatives, Direct File was created in-house at the IRS, in partnership with:

  • U.S. Digital Service (USDS)
  • General Services Administration (GSA)
  • Contractors like TrussWorks, Coforma, and ATI

This hybrid structure ensured agile execution while maintaining strong public stewardship.

🔒 Security Without Obscurity

Despite being open source, Direct File excludes any code that touches:

  • Personally Identifiable Information (PII)
  • Federal Tax Information (FTI)
  • Sensitive But Unclassified (SBU) data
  • National Security Systems (NSS) code

This reflects a disciplined balance between transparency and trust—one that more government software projects should emulate.

📜 Legal Framework

Direct File is anchored in a suite of progressive digital policies:

  • Source Code Harmonization And Reuse in IT Act of 2024
  • Federal Source Code Policy
  • Digital Government Strategy
  • E-Government Act of 2002
  • Clinger-Cohen Act of 1996

Together, these policies mandate that custom-developed government software should be shared and reused, not siloed.

💡 Why This Matters

Direct File represents a milestone for civic tech, open government, and digital service delivery:

✅ 1. Open Source, Real Impact
It’s not often we see real, working government platforms open to inspection and reuse. This invites contributions from civic technologists and helps other governments learn from U.S. innovation.

🧩 2. Designing for Complexity
Converting complex tax logic into user-friendly language—using a structured knowledge graph—is a pattern applicable well beyond taxes (think healthcare, benefits, or housing).

🛠️ 3. Engineering Innovation
The Fact Graph and modular backend architecture reflect best practices in modern backend design—resilient, flexible, and portable.

🔐 4. Trust and Privacy by Design
The selective code release shows how governments can be open while still securing sensitive systems.

🌐 5. Interoperability with State Systems
The state-api integration is especially forward-thinking. It could pave the way for smoother federal–state collaboration in everything from benefits to compliance.

🚀 Getting Started with the Code

Want to explore it locally? You’ll need:

  • Java
  • Scala
  • Maven
  • SBT
  • Coursier
  • Docker

Then: 1. Clone the repo

2. Run:

docker compose up -d --build

3. Navigate to /direct-file/df-client

4. Run:

npm run start

Frontend runs at http://localhost:3000

Final Thoughts

Direct File shows that government software doesn’t have to be clunky, slow, or hidden. With the right talent and commitment, it can be modern, secure, and open.

This project is not just about taxes—it’s about showing the public sector what’s possible when we build with purpose and publish with pride.

📎 Explore the repo (once public): github.com/irs-directfile

📩 Want to build something similar? Contact me

by dilanshah

AI Market Dynamics: Consolidation, Opportunity, and Innovation

The latest episode of “No Priors” featuring Sarah and Elad delves deeply into the current state of the AI market, revealing intriguing trends and untapped opportunities.

Consolidation and Expansion in AI

The AI market is undergoing notable consolidation in specialized sectors such as Large Language Models (LLMs), healthcare, and coding. These sectors are primarily driven by proprietary data sources, effective distribution channels, and widespread user adoption.

However, new entry points are emerging rapidly through open-source innovations like Microsoft’s Copilot and CodeStroll. These initiatives highlight a significant shift toward democratizing access to advanced AI capabilities. Yet, their ultimate success will heavily depend on their scalability and the consistent quality of outputs.

Meanwhile, markets for sales automation, productivity tools, and financial analytics remain largely fragmented without clear market leaders. This lack of dominance creates substantial room for innovation, competition, and investment, marking these sectors as particularly promising for entrepreneurs and innovators.

Intersection of Biotech and AI: Uncharted Opportunities

AI continues to unlock groundbreaking possibilities in biotechnology. Fertility treatments, stem cell differentiation, and egg maturation stand out as under-explored areas ripe with substantial commercial potential. Despite the enormous promise, many innovations remain undervalued and underfunded.

Conversely, groundbreaking research in muscle rejuvenation, tooth regeneration, and dental gene therapies faces significant hurdles due to their perceived low commercial appeal and associated developmental barriers. These scientific advancements await commercial champions willing to address these challenges and unlock their potential.

Tackling AI Development Challenges

The concept of building an “AI world model” encapsulates numerous open-ended research questions and challenges. Currently, scaling model size and the volume of training data remains fundamental for enhancing knowledge acquisition and pattern recognition in AI systems.

Reinforcement learning, despite its promise, struggles with challenges like adaptability and overfitting. To overcome these limitations, there’s a critical need to develop universal training environments and improved mechanisms for capturing and utilizing trace data effectively.

Novel AI Approaches

Innovative approaches such as evolved systems and self-selecting systems are pushing boundaries in AI development. These methodologies often yield superior outcomes by navigating search spaces through unconventional strategies, as evidenced by recent successes in molecular evolution experiments and advanced protein design.

By continually exploring and embracing such novel methodologies, AI development is set to achieve breakthroughs previously thought unattainable.

For a deeper exploration of these topics, watch the full episode on YouTube.

by dilanshah

Rise of AI Development Environments

The rise of Cursor, Copilot + VSCode, Replit, and Qwen2.5 among others, have caused me to rethink my ways. Focus will still be key in discerning what to build.


AI development environments change the global technology conversation. They also influence the pace of hiring and team augmentation decisions.

Qwen2.5-Coder Open Source

Alibaba Group has released the Qwen2.5-Coder open-source model. Qwen2.5-Coder-32B-Instruct is currently the best-performing open-source code model (SOTA), matching the coding capabilities of GPT-4o. Qwen2.5-Coder offers six different model sizes: 0.5B, 1.5B, 3B, 7B, 14B, and 32B.

Project: https://qwenlm.github.io/blog/qwen2.5-coder-family/

Each size provides both Base and Instruct models. The Instruct model engages in direct dialogue. The Base model serves as a foundational model for developers to fine-tune.

Github: https://github.com/QwenLM/Qwen2.5-Coder

Huggingface: https://huggingface.co/collections/Qwen/qwen25-coder-66eaa22e6f99801bf65b0c2f
Additionally, it provides two scenarios, code assistants and Artifacts, for exploration.

Code Assistants: https://huggingface.co/spaces/Qwen/Qwen2.5-Coder-demo
Artifacts: https://huggingface.co/spaces/Qwen/Qwen2.5-Coder-Artifacts

by dilanshah

The Future of Augmented Reality Glasses

Meta

  • Five years ago, plans were announced to create AR glasses to blend the digital and physical worlds.
  • Orion, the latest AR glasses from Meta, aims to enhance presence, connectivity, and empowerment.
  • AR glasses offer unrestricted digital experiences with large holographic displays.
  • Contextual AI integration allows for proactive addressing of user needs.
  • Orion stands out for its lightweight design, indoor/outdoor versatility, and emphasis on interpersonal interactions.
  • Orion represents a significant advancement in AR glasses technology, combining wearability with advanced features.
  • The groundbreaking AR display in Orion offers immersive experiences and a wide field of view.
  • Orion’s unique design maintains a glasses-like appearance, allowing users to see others’ expressions.
  • Orion’s capabilities include smart assistant integration, hands-free communication, and immersive social experiences.
  • While not yet available to consumers, Orion serves as a polished product prototype for future AR glasses development.

OP: Introducing Orion, [Meta’s] First True Augmented Reality Glasses

Snap

Sophia Dominguez, the Director of AR Platform at Snap, discussed Snap Spectacles and the company’s AR initiatives at the Snap Lens Fest original post here.

  • Snap Spectacles can be connected to a battery pack for extended use beyond the standard 45 minutes, with a focus on B2B interactions that directly engage consumers.
  • There is a push for Snap to collaborate with various businesses, such as those in location-based entertainment or museums, to expand the Snap Spectacles ecosystem.
  • Sophia Dominguez has been involved in AR for over a decade, starting with Google Glass, and now oversees developers and partners creating lenses on Snapchat.
  • Snap’s approach to AR emphasizes personal self-expression as a catalyst for AR lenses, transitioning to world-facing AR lenses like those in Snap Spectacles.
  • Snap’s long-term vision is to make AR ubiquitous and profitable for developers, aiming to integrate digital objects seamlessly into the real world.
  • Snap’s focus on consumer-level AR use cases includes self-expression as a core feature, offering a variety of options for users to engage with AR content.
  • Snap’s AR platform also caters to enterprise and B2B applications, collaborating with stadiums, museums, and other businesses for unique AR experiences beyond consumer-facing lenses.
  • Snap’s technology, like Snapchat cam, is designed for venues to integrate into large screens or jumbotrons, focusing on consumer desires for virality and joy rather than just enterprise solutions.
  • The company aims to increase ubiquity by making lenses fun and approachable, partnering with entities like the Louvre to explore augmented reality possibilities in a consumer-friendly manner.
    HTC Vive has delved into location-based entertainment more than Meta, and Snap is prioritizing connected experiences, ensuring fast connectivity and optimizing for various use cases like museum activations.
  • Snap collaborates closely with developers, offering grants and support without strings attached to foster innovation in the augmented reality space, aiming to be the most developer-friendly platform globally.
  • Snap’s Spectacles have evolved over the years, from simple camera glasses to AR display developer kits, with the latest fifth generation focusing on wearability, developer excitement, and paving the way for consumer adoption.
  • The company has revamped Lens Studio to encompass mobile and Spectacles lenses, emphasizing ease of use and spatial experiences, aiming to create a seamless ecosystem for developers across different platforms.
  • Snap values feedback and collaboration with developers, striving to provide pathways for monetization and support for creators building on both mobile and Spectacles platforms.
  • Snap’s Spectacles offer a unique immersive experience, leveraging standalone capabilities and spatial interactions, aiming to enable emergent social dynamics and experiences not possible on other devices.
  • Developers are considering the length of time users spend on devices like Zoom calls or workouts, with a focus on creating a seamless experience for users on the go.
  • The new SnapOS manages a dual processing architecture for Spectacles, with Lens Studio being the primary pipeline for developers to create content for the device.
  • Snap is actively listening to developer feedback and working on enabling WebXR on Spectacles to support a variety of use cases and experiences.
  • The operating system for Spectacles includes features like connected lenses, hands and voice-based UI, and social elements out of the box to facilitate easier development.
  • The ultimate potential of spatial computing is envisioned as a way to break free from the limitations of screens, allowing for more natural interactions and connections in the real world.
  • Snap aims to empower developers to explore the possibilities of augmented reality and spatial computing, emphasizing ease of use and continuous improvement based on user feedback.

by dilanshah

Useful Resources for AI

Newsletters/blogs:
– TLDR AI (https://tldr.tech/ai) – Andrew Tan
– Ben’s Bites (https://lnkd.in/gNY8Dmme)
– The Information *paid subscription required (https://lnkd.in/gbkaFbvf)
– Last week in AI (https://lastweekin.ai/)
– Eric Newcomer (https://www.newcomer.co/)

Podcasts:
– No priors with Sarah Guo + Elad Gil (https://lnkd.in/g7Wmr6XT)
– All-in podcast – not AI specific but they talk a lot about it (https://lnkd.in/gH35UeUy)
– Lex Fridman (https://lnkd.in/gjw7zsWX)

Online courses:
– DeepLearning.ai by Andrew Ng – https://lnkd.in/gWcn5UTK

Institutional VC writing: 
– Sequoia (https://lnkd.in/g-cKpn8Y)
– A16z (https://lnkd.in/g6JxqwZA)
– Lightspeed (https://lnkd.in/gczzdEcd)
– Bessemer (https://www.bvp.com/ai)
– Radical Ventures (https://lnkd.in/guCe5Mnt); Rob Toews (https://lnkd.in/ggH8HfT8) and Ryan Shannon (https://lnkd.in/gRrBzePx)
– Madrona (https://lnkd.in/gy5D8yNG)

Industry Conferences:
– Databricks Data + AI Summit (https://lnkd.in/gF5QyXYv)
– Snowflake (https://lnkd.in/gavqzw65)
– Salesforce Dreamforce (https://lnkd.in/gJk4r58N)

Academic Conferences:
– NeurIPS (https://neurips.cc/)
– CVPR (https://cvpr.thecvf.com/)
– ICML (https://icml.cc/)
– ICLR (https://iclr.cc/)

Books:
– Genius makers, by Cade Metz (https://lnkd.in/gr_78MB9)
– A Brief History of Intelligence, by Max Bennett (https://lnkd.in/g2uCrPzS)
– The worlds I see, by Fei-Fei Li (https://lnkd.in/gY8Qsvis)
– Chip Wars, by Chris Miller (https://lnkd.in/g6ZAZSCG)

The original author of this post was Kelvin Mu on Linkedin.