ManiKumar Jami

Built India’s first voice AI, blockchain edtech, auctions on WhatsApp. AI product leader, entrepreneur, 3X founder. behavioural psychology and legal tech

[ManiKumar Jami @manikumarjami]

@ncbn Side of the people ? Daaamnnn, the tweet started with a vision.

[Ram Gopal Varma @RGVzoomin]

In effect what you mean is , in a democracy, a leader is given a responsible position by people who trusted him to take care of their issues ..He’s not been crowned as a king 🙏

[ManiKumar Jami @manikumarjami]

@PawanKalyan not sure about your academics, but wanted to see your scores in a lie detector test.

But I appreciate your confidence while talking baseless shit and propaganda politics.

Laddu is over, sanathani is over, now AP-TG conflicts and YSR.

UFFFFFF. TRULY A GREAT ACTOR

[ManiKumar Jami @manikumarjami]

Something that questions whatever we are doing everyday and every where...

https://t.co/8lqyMNQsYi

[ManiKumar Jami @manikumarjami]

Claude for architecture, GPT-4o for boilerplate, and Qwen 3.5 for iteration loops. The trick isn't picking one -it's knowing which task fits which model's cost to quality ratio. Bookmark this.

https://t.co/bALDQQdoZR

[ManiKumar Jami @manikumarjami]

Unpopular opinion: the devs writing the best code with LLMs aren't prompting better. They're just reading the output more skeptically than you are.

[ManiKumar Jami @manikumarjami]

Most research papers that go viral on social media are already outdated… about as outdated as benchmarking today’s tech against its predecessors.

By the time it trends, the world has already moved on.

@OpenAI @AnthropicAI @GeminiApp

[ManiKumar Jami @manikumarjami]

Every new car will have federal surveillance tech by 2027. You won't be able to opt out.

[ManiKumar Jami @manikumarjami]

Everyone's obsessing over which AI model is "best." The real story is how the architecture underneath is quietly forking into completely different species.

Sebastian Raschka just published something that deserves way more attention than it's getting: a visual gallery tracing how large language model architectures have evolved from 2017 to today. On the surface, it looks like a helpful reference guide. But if you actually study the diagrams, you'll notice something the benchmark leaderboards completely miss. We're not watching incremental improvements to one design. We're watching the emergence of fundamentally different computational organisms, each optimized for different survival strategies.

The original Transformer architecture from Google's 2017 "Attention Is All You Need" paper had a clean encoder-decoder structure. Both halves. Balanced. That design dominated early models like BERT and the original T5. But look at what happened next. GPT took the decoder-only path. BERT went encoder-only. And now, in 2024 and 2025, we're seeing architectures that would be unrecognizable to someone who stopped paying attention in 2020. Raschka's gallery documents this divergence with precision, and the implications matter more than any single benchmark score ever could.

The popular narrative goes something like this: OpenAI builds big models, competitors build slightly different big models, and the race is about who can scale parameters and training data fastest. That's the story most people believe. It's also increasingly wrong.

The evidence tells a more complicated story. Raschka's diagrams show that modern architectures like Llama, Mistral, and Gemma have converged on decoder-only designs, but with critical internal differences. Llama uses RMSNorm and RoPE positional embeddings. The original GPT used learned positional embeddings and LayerNorm. These sound like minor technical details until you realize they change how models handle long contexts, how efficiently they train, and how they can be adapted for specific tasks.

Here's where it gets interesting. Raschka highlights architectural innovations that don't make headlines but matter enormously for practitioners. Grouped Query Attention, which Llama 2 introduced, reduces memory requirements during inference by sharing key-value heads across query heads. This isn't a parameter count war anymore. It's an efficiency war. Meta's Llama 3 and Llama 4 use this technique to serve models that would otherwise be impractical at scale.

Meanwhile, Google's Gemma architecture takes a different approach, optimizing for deployment flexibility. Mistral's models emphasize sliding window attention for handling longer sequences without quadratic memory growth. Each represents a genuine architectural bet about what constraints will matter most in production.

The original Transformer from 2017 had roughly 65 million parameters in its base configuration. Today's frontier models like GPT-4 are estimated at over 1 trillion parameters. But Raschka's gallery shows that the architectural distance traveled is actually more significant than the parameter growth. You can't just "scale up" a 2017 Transformer and get GPT-4. The internal machinery is different.

Here's what the conventional wisdom gets wrong: we keep talking about large language models as if they're one category of thing, like "cars" or "smartphones." But studying Raschka's architectural evolution suggests a better analogy might be biological speciation. Just as mammals and birds both descended from common ancestors but evolved radically different solutions to flight, these model architectures are diverging toward different ecological niches.

Decoder-only models like GPT and Llama have won the general-purpose text generation niche. But encoder-decoder models like T5 and FLAN still dominate certain translation and summarization tasks. And we're now seeing hybrid approaches, like retrieval-augmented architectures, that graft external memory systems onto base models.

I might be wrong about this, but I think the "which model is best" framing will look as outdated in two years as asking "which animal is best" looks to a biologist. The answer is always "best at what, in what environment?"

The part most people miss if they only follow benchmark scores: architectural choices made years ago constrain what's possible today. RoPE embeddings enable better long-context performance but require different fine-tuning approaches. Grouped Query Attention saves inference cost but changes how the model allocates representational capacity. These aren't neutral engineering decisions. They're bets about the future.

If you're a developer, this matters because the model you build on today will inherit its architecture's limitations. If you're an investor, this matters because you're not betting on one race. You're betting on which of several parallel evolutionary paths leads to the most valuable applications.

By the end of 2026, I expect we'll see at least three distinct "species" of production large language models that are no longer meaningfully comparable on shared benchmarks. One optimized for long-context reasoning with architectures descended from Llama's RoPE lineage. One optimized for multimodal integration with architectures that look more like Gemini's hybrid approach. And one optimized for edge deployment with aggressive efficiency techniques we're only beginning to see in models like Mistral and Phi.

The company that wins isn't the one with the highest score on a benchmark designed for last year's architectures. It's the one that correctly identifies which niche will be largest and most defensible, then optimizes their architecture for that specific environment.

The losers will be companies still treating "large language model" as a monolithic category and making architecture-agnostic bets.

What would prove me wrong? If scaling laws continue to dominate and architectural innovations plateau, then sheer compute and data will still be the primary differentiators. But Raschka's gallery suggests the opposite trend. The pace of architectural innovation is accelerating, not slowing.

If I'm right about this architectural speciation, what happens when enterprises need to choose not just a vendor, but a computational species to build their future on?

[ManiKumar Jami @manikumarjami]

🧵 (4/4) Knuth was 40 years early. The best interface for agent-generated code might be the one humans refused to write themselves. Anyone building agent tooling that treats prose and code as first-class outputs together? Drop examples below.

[ManiKumar Jami @manikumarjami]

🧵 (1/4) Literate programming flopped for humans. It might be perfect for AI agents.

[ManiKumar Jami @manikumarjami]

🧵 (3/4) Here's what most agent builders miss: when an AI explains its reasoning inline with code, you get automatic documentation, easier debugging, and a reviewable chain of thought. The "overhead" becomes the feature - agents that narrate their work are agents you can trust.

[ManiKumar Jami @manikumarjami]

1,250 dead in Haiti from drone strikes and AI Twitter's debating benchmark scores.

Autonomous weapons aren't a future ethics problem - they're deployed now. Where's the urgency?

[ManiKumar Jami @manikumarjami]

The future of coding isn't more comments. It's prose-first programming where humans write the story and AI writes the code.

Donald Knuth introduced literate programming in 1984. The idea was radical: instead of writing code with occasional comments, you'd write a document explaining your thinking, with code woven throughout. The program becomes a narrative. Knuth built TeX this way, and WEB, his literate programming system, produced some of the most maintainable software ever written.

Almost nobody adopted it. The tooling was clunky. Developers found it slower than just coding. By the 1990s, literate programming became a footnote in computer science history, something academics admired but practitioners ignored.

Now AI coding agents are everywhere. Cursor, GitHub Copilot, Replit Agent, Devin. Anthropic reports Claude writes 30% of its own codebase. Cognition's Devin can work autonomously for hours on complex tasks. These agents don't just autocomplete - they reason, plan, and execute multi-step implementations.

Here's the tension nobody's addressing: we're giving these agents free rein over codebases they don't understand. They can read syntax perfectly but miss intent entirely. The gap between what you meant and what the agent builds is widening, not shrinking.

The evidence suggests we've been solving the wrong problem for forty years. Knuth wasn't too early. He was building for a reader that didn't exist yet.

Consider what actually happens when you use an AI coding agent today. You write a prompt. The agent generates code. You review it, find issues, prompt again. Repeat. Anthropic's research on Claude's self-coding shows the model performs best when given extensive context about architectural decisions, constraints, and reasoning. When that context is missing, the agent makes locally reasonable choices that create global inconsistency.

Simon Willison, one of the most prolific AI-assisted developers working today, has documented this pattern extensively. His workflow involves writing detailed specifications before asking Claude to implement anything. He's essentially doing literate programming in reverse - writing the narrative first, then having AI fill in the code. His posts show projects where the specification document is three times longer than the final implementation.

Jupyter notebooks offer a glimpse of this future. They blend markdown explanation with executable code blocks, and they've become the standard in data science precisely because the narrative matters as much as the execution. A 2023 GitHub analysis found over 10 million public Jupyter notebooks, up from 2.5 million in 2018. Data scientists discovered what Knuth knew: when you're exploring complex problems, prose helps you think.

The pattern repeats in unexpected places. Notion's engineering team writes Architecture Decision Records for every significant choice. Stripe's API documentation reads like a tutorial because they learned developers understand systems through stories, not specifications. These aren't coding practices - they're writing practices that happen to involve code.

If you've watched an AI agent struggle with a codebase, you know the failure mode. It's not that the agent can't write working functions. It's that the agent doesn't know why a particular approach was chosen over alternatives. It doesn't know the team tried a different architecture in 2022 and it failed. It doesn't know the weird edge case that makes the obvious solution wrong.

Comments don't capture this. READMEs don't capture this. What captures it is narrative - the kind of connected, contextual explanation Knuth designed literate programming to produce.

I changed my mind on this recently. I used to think better prompting would solve the context problem. Give the agent more examples, better instructions, clearer constraints. But after watching agents hallucinate confidently on projects with excellent documentation, I realized the format matters as much as the content. Agents need prose they can follow like a story, not scattered comments they have to piece together.

The conventional wisdom says we should make code self-documenting. Clean variable names, small functions, obvious structure. That worked when humans were the only readers. But AI agents parse code differently than humans do. They're trained on natural language. They reason better when given natural language context. The "self-documenting code" ideal optimizes for the wrong reader.

Think of it like giving directions. You could hand someone a street map with optimal routes highlighted. Or you could say "head toward the old church, turn left at the coffee shop where we met last year, and it's the blue house with the garden out front." The second version contains landmarks, history, reasoning. It's less precise but more robust to confusion.

Literate programming treats code like the second kind of directions. The prose provides landmarks. The code provides precision. Together, they give an AI agent something neither can provide alone.

Here's my prediction: by the end of 2026, at least one major development framework will ship with literate programming as a first-class feature, not a plugin or afterthought. The tools will look different from Knuth's WEB - probably closer to Jupyter notebooks with better version control and agent integration. But the core idea will be the same: prose-first, code-second.

The winners will be teams that treat their codebase as a document to be read, not just executed. The losers will be teams that assume AI agents can infer intent from syntax alone. I might be wrong about the timeline - maybe it takes until 2028. But I'm confident about the direction.

The condition that would prove me wrong: if AI agents develop genuine long-term memory and reasoning about codebases without explicit narrative context. If they can somehow absorb institutional knowledge from code alone. Current architectures don't suggest this is coming soon, but the field moves fast.

Knuth built literate programming for a human reader who never showed up in numbers. Forty years later, we're building readers that might finally appreciate what he made.

What's your experience with AI agents and codebase context? I'm curious whether anyone's already experimenting with prose-heavy documentation specifically for agent consumption.

[ManiKumar Jami @manikumarjami]

Someone was asking that if Claude Code can now write code, run the app, debug errors, review pull requests, and even fix bugs automatically, what exactly is left for junior developers to do?

IYKYK

[ManiKumar Jami @manikumarjami]

Art isn't about output.

Everyone's obsessed with AI generating music. The real flex is AI sorting algorithms making the Amen break sound broken in new ways.

[ManiKumar Jami @manikumarjami]

Thomas Selfridge died in 1908 testing Orville Wright's plane. Four years later, commercial aviation launched anyway.

AI safety incidents won't slow deployment—they'll just shift liability.

[ManiKumar Jami @manikumarjami]

BMW just announced humanoid robots will work alongside humans in its German factories by the end of 2025. The buried detail? This isn't a pilot. It's a procurement decision.

The press release dropped quietly on June 4, 2025. BMW Group confirmed it will deploy Figure's humanoid robots at its Spartanburg, South Carolina plant - the company's largest globally - and extend deployment to its home production network in Germany within the same year. Spartanburg alone produces over 1,500 vehicles daily. The German plants include the Munich headquarters facility, where BMW has manufactured vehicles since 1922.

Here's what makes this different from the robot demonstrations you've seen before. BMW isn't running a lab experiment. They're not filming a choreographed demo. They signed what Figure describes as a "commercial agreement" - meaning money changes hands based on robots actually doing work. The specific initial task is logistics: moving parts containers from storage to the production line. Simple, yes. But it's the exact kind of repetitive, physically demanding work that manufacturing has struggled to staff for years.

Milan Nedeljković, BMW's Board Member for Production, said something worth paying attention to: they're treating humanoid robots as "a new type of versatile automation" that will supplement their existing workforce. Not replace. Supplement. That framing matters because BMW, like most German automakers, operates under strict labor agreements. You don't casually announce workforce changes in Germany without careful positioning.

The tension is obvious. If this works at scale, what exactly does "supplement" mean in three years?

I've been tracking humanoid robotics deployments for two years now, and what struck me about this announcement isn't what BMW said - it's what they didn't have to say anymore. In 2023, when Figure first emerged, the question was whether these robots could even stand up reliably in real environments. By early 2024, the question shifted to whether they could do useful tasks. Now, in mid-2025, we've jumped straight to commercial contracts with one of the world's most demanding manufacturing operations.

The detail most people skipped: BMW explicitly mentioned that Figure robots will operate in unstructured environments. That's industry code for "we're not redesigning the factory floor around the robot." Traditional industrial automation requires cages, fixed paths, and highly controlled conditions. A humanoid that can navigate existing spaces, handle variable container positions, and work near humans represents a fundamentally different deployment model.

Figure's robots use a vision-based AI system that BMW says enables them to "adapt dynamically to changing conditions." I've seen similar claims from a dozen robotics companies. What's different here is BMW put their manufacturing reputation on the line by making it public. German automakers don't announce production technology partnerships for marketing buzz. They announce them when the technology is ready for the floor.

The numbers tell a story. BMW employs roughly 155,000 people globally, with over 80,000 in Germany alone. Their Spartanburg facility has about 11,000 workers producing X3, X5, X7, and XM models. Figure has raised over $750 million, including investment from BMW's venture arm. The commercial agreement covers 2025 deployment, which means we'll know within 18 months whether this is real or another robotics overpromise.

Here's what I think the conventional wisdom gets wrong: most coverage frames this as "robots taking jobs" or "automation finally arriving." Both framings miss the actual story.

The real story is that the humanoid form factor just got validated by the world's most quality-obsessed manufacturing culture. BMW's German plants operate under the "zero defects" philosophy. If you've ever been inside one, you know they're almost absurdly precise. Introducing humanoid robots into that environment isn't a statement about labor costs. It's a statement about capability maturity.

I changed my mind on humanoid robots about eight months ago. I used to think the human form was an inefficient engineering choice - why build legs when wheels work fine? The answer, I now believe, is infrastructure compatibility. Every factory, warehouse, and workplace on Earth was designed for humans. Stairs, handles, switches, aisles. A humanoid robot can theoretically operate in any human workspace without modification. That's not a small advantage. That's potentially the entire advantage.

Think of it like this: building a specialized robot for every task is like writing custom software for every computer. Building one humanoid platform that can learn multiple tasks is like having an operating system. BMW isn't betting on one task. They're betting on a platform.

The screenshot-worthy observation: BMW didn't need to do this publicly. They could have run pilots quietly for years. The fact that they announced a commercial deployment to their German plants, in the same breath as the US deployment, suggests they see a competitive window. If your rivals are three years behind on humanoid integration, you don't hide your progress. You signal.

By the end of 2026, I expect at least three more major European automakers will announce similar partnerships with humanoid robotics companies. Volkswagen, Mercedes-Benz, and Stellantis are all facing the same workforce dynamics: aging labor pools, declining interest in factory work among younger workers, and relentless pressure to match Asian manufacturing efficiency.

The condition that would prove me wrong: if BMW's German deployment gets delayed past Q2 2026, it likely means the technology hit practical limits that controlled US environments didn't reveal. German works councils have significant power to slow or block automation that threatens jobs. If IG Metall, Germany's powerful metalworkers' union, pushes back hard, the "supplement" framing might not hold.

Who wins in this scenario? Figure obviously, but also any company building humanoid-compatible AI systems. NVIDIA's robotics simulation platforms, for instance, become more valuable if humanoids become standard manufacturing equipment. Who loses? Specialized automation vendors who've built business models around single-purpose machines. If one robot can do ten tasks, you don't need ten robots.

The longer play is even more interesting. If humanoid robots prove viable in automotive manufacturing, the second market won't be other factories. It'll be eldercare. Germany has one of the oldest populations in Europe. BMW figuring out human-robot collaboration in controlled factory environments could be the training ground for robots that eventually work in homes.

What am I missing here? I'm focused on the manufacturing angle because that's what the announcement covers. But BMW also has massive R&D operations. If you work in automotive robotics or have insight into how German works councils are actually responding to this news, I'd genuinely change my view based on that ground-level context.

[ManiKumar Jami @manikumarjami]

Everyone's debating whether Apple backdoors exist. The real story is that the front door was already wide open.

The headlines this week focused on the theoretical threat: could governments force Apple to build iPhone vulnerabilities? Meanwhile, a very real hacking toolkit - one reportedly used by US federal agencies - has apparently spread to foreign intelligence services and criminal organizations. The debate about what might happen someday is obscuring what's already happened.

Grayshift, an Atlanta-based company founded in 2016 by former Apple engineers, built a device called GrayKey specifically designed to crack iPhone passcodes. The company has sold these tools to US law enforcement and federal agencies, including Immigration and Customs Enforcement and the Secret Service. For years, GrayKey represented one of the most potent iPhone-hacking capabilities outside of nation-state intelligence services. The device could bypass Apple's security measures through methods the company has kept closely guarded.

Now, according to reporting from Wired, there's evidence suggesting these capabilities - or tools derived from them - have ended up in hands they were never meant to reach. We're talking about foreign governments and organized criminal groups who've obtained access to iPhone-cracking technology that was supposed to be controlled, regulated, and limited to legitimate law enforcement purposes.

The stakes here extend beyond privacy advocates' concerns. If you're a journalist, activist, business executive, or anyone who relies on iPhone security while traveling internationally, the threat model just changed. Your device's security isn't just about Apple's engineering anymore. It's about supply chain security for the tools designed to defeat that engineering.

The popular narrative around iPhone security goes something like this: Apple builds strong encryption, governments complain they can't access criminal evidence, and the standoff continues indefinitely. Privacy wins by default because the math behind encryption is unbreakable.

This framing misses something critical. The security industry has spent over a decade building workarounds. Grayshift is just one player. Cellebrite, an Israeli company, offers similar capabilities and has contracts with law enforcement agencies in over 100 countries. NSO Group's Pegasus spyware, which exploits iPhone vulnerabilities, has been documented targeting journalists and dissidents in countries including Saudi Arabia, Mexico, and the UAE.

Here's where the evidence gets uncomfortable. Grayshift's tools were designed with safeguards - they're supposed to be sold only to vetted agencies, operated on-site, and tracked. But the company's core technology relies on exploiting iOS vulnerabilities. Once a vulnerability is discovered and weaponized, that knowledge becomes a transferable asset. A former employee leaves. A device gets stolen. A foreign agency reverse-engineers the approach.

The Wired report indicates that techniques consistent with Graykey capabilities have appeared in forensic operations conducted by governments that were never authorized customers. Criminal organizations have demonstrated similar iPhone-cracking abilities in contexts that suggest access to commercial-grade tools, not just nation-state resources.

Consider the math. Grayshift reportedly charges between $15,000 and $30,000 for device licenses. That's pocket change for a drug cartel generating billions annually. It's trivial for a foreign intelligence service. The economic barrier that was supposed to limit these tools to legitimate users is essentially meaningless against motivated adversaries.

Apple patches vulnerabilities when they're discovered, but the cat-and-mouse dynamic means there's always a window. Grayshift and Cellebrite employ teams specifically to find new exploits. Apple's iOS 17 introduced Lockdown Mode specifically to counter sophisticated attacks, but adoption remains limited to users who know they're at risk.

The conventional wisdom gets this wrong in a fundamental way. We've been treating government backdoor mandates as the primary threat to mobile security. The encryption debate - should Apple be forced to build access for law enforcement? - has consumed years of policy discussion, congressional hearings, and court battles.

But the threat that actually materialized wasn't a mandated backdoor. It was a commercial front door, sold legally, that escaped its intended constraints. You don't need to force Apple to weaken encryption when you can buy tools that exploit implementation flaws in the operating system itself.

I might be wrong about the scale here, but I don't think I'm wrong about the direction. The proliferation pattern we've seen with Pegasus - starting with premium government customers, then spreading to problematic regimes, then showing up in targeted attacks on civilians - appears to be repeating with physical device-cracking tools.

The security community has focused intensely on zero-click exploits and remote attacks because they're technically elegant and scary. The boring reality is that physical device access combined with commercial cracking tools represents a more immediate risk for most people. If someone can get their hands on your phone for a few hours, the encryption protecting your data may no longer be the relevant security boundary.

This also explains why Apple's recent hardware security improvements - the Secure Enclave, USB Restricted Mode, stolen device protection - matter more than software updates alone. Apple is essentially racing against an industry that exists specifically to defeat their previous generation's protections.

Within 18 months, I expect we'll see documented cases of GrayKey-derived capabilities being used by at least three additional foreign intelligence services beyond the current known customer base. The commercial mobile forensics market is too lucrative and the technology too transferable for containment to hold.

Cellebrite will face increased pressure to implement technical restrictions that prevent device operation outside authorized jurisdictions. They'll resist, citing customer needs. Eventually a high-profile case - probably involving a journalist or opposition figure in a US-allied nation - will force regulatory action.

Apple will respond by accelerating hardware-based protections that make physical device attacks significantly harder. I'd predict iOS 19 or 20 will include boot-level security changes that invalidate current-generation forensic tools entirely, forcing another expensive development cycle for Grayshift and competitors.

The loser in this scenario is anyone who assumed iPhone security was binary - either Apple built a backdoor or they didn't. The winner is whoever correctly prices in the reality that security tools, like weapons, proliferate according to economic incentives rather than policy intentions.

The condition that would prove me wrong: if Grayshift and Cellebrite implement verifiable technical controls that genuinely prevent tool operation by unauthorized parties, and those controls survive independent security audits. I'm not holding my breath.

If I'm right about this trajectory, what happens when the next generation of AI-powered forensic tools emerges - ones that can crack devices faster and cheaper than anything we've seen? The proliferation problem we're watching unfold with GrayKey is just the preview.