Raavi.dev

AI Software Engineer & Solo Founder.
Building intelligent software for real-world problems.
Creator of BiteFix — iOS clean grocery scanner 👇

[Raavi.dev @RSP_ER]

Stop reading.
Start scanning.
Get instant insights.

BiteFix is an evergreen pocket utility for grocery shoppers and snack buyers to make instant, clean eating decisions with zero friction.

Check out the IOS app yourself: https://t.co/u4qZpgLftP

[Raavi.dev @RSP_ER]

Sam Altman literally renamed his X account to "Ram Altman" after claiming your gaming PC is wasting the world's memory supply.

The AI compute bottleneck has officially moved from GPU silicon to DRAM allocation.

Why:
Every frontier model training run and inference cluster is starved for high-bandwidth memory (HBM) and DDR5 DRAM.

During a recent interview, Sam Altman openly lamented that millions of gigabytes of high-performance RAM sit completely idle inside consumer desktops while AI data centers face severe allocation queues.

Hardware creators and developers fired back immediately:

- Memory allocation is not a shared utility pool.

Consumer DDR5 and data-center HBM3e serve completely different memory bus architectures and bandwidth profiles.

- AI clusters consuming hundreds of megawatts are feeling the squeeze of high-memory pricing and supply constraints.

- Shifting blame onto gaming rigs and local workstations ignores the reality of datacenter infrastructure planning.

Whether it was a Silicon Valley joke or a candid frustration over memory supply chains, one thing is clear: memory bandwidth is the next trillion-dollar AI bottleneck.

[Raavi.dev @RSP_ER]

62% of your AI coding agent bill is not computation.

It is your model paying full price to reread what you already sent it.

Here is the exact architectural trap burning cloud credits, and the single primitive that cuts it in half:

THE INVISIBLE RESEND TRAP

Modern LLMs keep zero state between steps.

When an AI agent runs a multi-step task - reading files, checking compiler errors, testing hypotheses, and editing code:

• Step 1: You send standing system instructions, tool schemas, and the user prompt. (Cost: 2,000 tokens)

• Step 2: To make one minor file edit, the agent resends 100% of Step 1, plus tool outputs and diffs. (Cost: 5,000 tokens)

• Step 3: To run unit tests, it resends all of Step 1 and Step 2 again. (Cost: 12,000 tokens)

Only the newest instruction is actually new.
The rest is an identical massive block of static text resent at full input pricing on every single iteration.

WHY STANDARD COMPACTION FAILS

When you wait for an agent to run 20 turns before compacting context, the financial damage is already done.

In a loop of N turns:

Total Billed Input ≈ (N² / 2) × Average Context Added

Stanford and Gartner benchmark studies confirmed that across typical enterprise agent sessions, 62% of total token fees were spent reprocessing unchanged context blocks.

THE FIX: PROMPT CACHING HEADERS

Leading frontier providers (Anthropic, OpenAI, DeepSeek) now support KV-cache retention on server-side model nodes.

When you pin static tokens at the top of your request:

1. System instructions, project architecture maps, and tool definitions are cached on the provider's GPU memory.

2. Subsequent turns read cached prefixes with up to 90% cost discounts and 80% lower latency.

3. You only pay full rate for newly appended turns and tool responses.

If your orchestration framework is not utilizing prompt caching, you are leaving 25% to 50% of your operational budget on the table.

[Raavi.dev @RSP_ER]

Every frontier AI model scored 0% on ProgramBench at launch.

John Yang - the researcher behind SWE-bench, SWE-smith, and CodeClash- just dropped the benchmark that exposed the real ceiling of autonomous coding agents.

Here is why it broke every model on earth:

SWE-bench measures how well an agent can patch an existing codebase given a GitHub issue and regression tests. It’s localized bug-fixing.

ProgramBench is the exact opposite: greenfield reverse-engineering from scratch.

Models are given only:
1. A compiled binary
2. User-facing documentation

From those two alone, the agent must architect, implement, and compile a complete software program that achieves identical behavioral equivalence against dynamic test suites (inspired by Nicholas Carlini’s work on C compilers).

When it launched last week:
Every single frontier model—Claude 3.5 Sonnet, GPT-4o, Gemini 1.5 Pro—hit a flat 0% pass rate.

Amazing!

[Raavi.dev @RSP_ER]

Why does almost every AI video look completely fake?

If you look closely at 99% of AI clips scrolling your feed, you spot the tells immediately:

• Skin that looks like airbrushed plastic
• Glassy, dead eyes with no emotion
• Faces that subtly morph into a different person mid-sentence
• Perfectly linear camera pans that scream "computer animation"

Making an AI video look 100% human and realistic isn't about expensive tools.

it comes down to 3 simple rules:

1. The 3-Photo Identity Lock
Instead of pasting one random selfie, feed the system 3 angles: front view, side profile, and an expressive close-up in flat natural daylight. This locks bone structure so your face never changes shape mid-take.

2. The Raw First-Frame Anchor
Never start generating video from plain text. Create one single still image first with unpolished lighting (hard flash, raw daylight, subtle sensor grain). The video generator borrows its color science, contrast, and mood directly from that one master frame.

3. Handheld Camera Physics (The Imperfection Budget)
Real humans don't film with mechanical robot arms. Add a gentle handheld sway, let the autofocus hunt for a fraction of a second, and shoot in one continuous 15-second take with zero cuts.

Real life is messy, textured, and imperfect. When you stop chasing perfection, your AI video finally looks real.

[Raavi.dev @RSP_ER]

"The junior developer is dead. English is the new programming language."

Everyone is repeating this soundbite on X.

Here is what is really happening to software engineering:

Typing syntax was never the core job.

Writing boilerplate loops, plumbing APIs, and formatting CRUD operations was just the manual labor tax we paid because computers couldn't interpret intent.

Now we have three completely different software paradigms operating at the same time:

1. Software 1.0 (Explicit Code): Written in C++, Python, Rust. Deterministic, fast, exact.

2. Software 2.0 (Neural Net Weights): Learned from data. Pattern recognition, vision, perception.

3. Software 3.0 (Prompting LLMs): Programmed in natural language. Orchestrating reasoning, high-level logic, and agents.

If your entire skill set as a junior dev was typing boilerplate that a 7B model generates in 200ms, yes that specific job is gone.

But if you understand system architecture, data flow, failure modes, and when to fluidly transition between 1.0, 2.0, and 3.0:

You don’t have fewer engineering problems to solve.

You just have 100x more leverage than any engineer in history.

Stop mourning syntax.
Start mastering architecture.

[Raavi.dev @RSP_ER]

@Phinneas_me Even the names of the tasks you mentioned are scary enough to understand that building a business is not just prompts, right.

[Raavi.dev @RSP_ER]

Every founder on X is chasing the "One-Person Unicorn."

The thesis sounds irresistible:

Prompt an autonomous agent.

Generate a full-stack SaaS in 48 hours.
Let AI handle the code, tests, and deployment while you sit back and collect recurring revenue.

Here is the reality check:

AI made software generation essentially free. Which means the barrier to shipping code dropped to zero.

When writing software costs nothing, software itself ceases to be a competitive moat.

“You don't win a startup by having expertise in startups or obsessing over mechanics. You win by having obsessive, unreplicable expertise in your users.”

Mark Zuckerberg didn’t scale Facebook because he was a master engineer or venture wizard.

Facebook was originally incorporated as a Florida LLC. He won because he understood the social dynamics of college students better than anyone on earth.

If your entire startup moat is "I can prompt Cursor or Claude Code faster than you," you don't have a business you have a weekend demo that 50 other builders will duplicate by Monday morning.

The future doesn't belong to the person who generates code the fastest.

It belongs to the founder who solves distribution, commands customer trust, and builds something people actually refuse to live without.

Focus on the users. The code is just plumbing.

[Raavi.dev @RSP_ER]

@NivasT49781 Yes that’s what drives traffic for all of them selling hopes to everyone that they can create a business by some prompts.

[Raavi.dev @RSP_ER]

@sahildarz Yes bro I agree, it takes countless nights of back and forth testing , fixing, maintaining.

N I can tell you it’s a nightmare when perfecting even onboarding design part. 😂😂

[Raavi.dev @RSP_ER]

Most developers are setting fire to 70% of their LLM context window before typing a single prompt.

And it’s all coming from how we’re implementing MCP.

Here is the dirty secret of the MCP boom:

To give an LLM a tool via MCP, its entire JSON schema, parameter tree, and docstring must be injected directly into the system prompt.

Geoffrey Huntley (@ghuntley) recently audited the numbers on the official GitHub MCP server:
• 93 tools exposed
• ~55,000 tokens consumed upfront purely for tool definitions

If you connect GitHub, Postgres, and Slack MCP servers in a standard 200k context window:

1. You burn 120,000+ tokens on initialization.

2. You pay input token pricing on those schemas on every single turn.

3. The LLM’s attention mechanism degrades drastically (“needle in a haystack” syndrome) trying to distinguish between 200 granular function signatures.

We solved this problem with standard CLIs.

If you give an autonomous agent access to a bash terminal and the GitHub CLI (`gh`):

• Upfront token cost: 0 tokens.

• Pre-training familiarity: Frontier models already memorized standard CLI flags.

• Dynamic discovery: If the model forgets a flag, it runs `gh issue create --help` on demand.

Dumping an entire REST API into an LLM's working memory isn't "agent architecture"—it's schema bloat.

Progressive tool discovery and lightweight CLIs win every time.

[Raavi.dev @RSP_ER]

"Please don't execute dangerous shell commands" is not a security architecture.

Prompt guardrails are vibes.
Docker containers are shared-kernel theater.

The moment you hand autonomous AI agents CLI access, package managers, and web scrapers, you are not dealing with prompt injection—you are dealing with lateral network movement, ambient IMDS credential theft, and container escapes.

In production, autonomous runtime containment has only two real answers:
1. Ephemeral microVMs (Firecracker / non-shared kernels)
2. Deterministic ABAC on every tool execution

Everything else is just waiting for a breach report.

[Raavi.dev @RSP_ER]

@pcstyle53 Bro where were till now , what a great web app you just showed us.
Thanks

[Raavi.dev @RSP_ER]

https://t.co/UtcqObQaDb

[Raavi.dev @RSP_ER]

Gemini 4 Argon just blew my mind.

I was watching a movie on streaming site it today started opening a new ad tab with almost every click.

Got frustrated, explained the problem to Gemini in 2 lines, and…

It built me a Chrome extension.
With an icon.

Amazing part is that
It has no UI
No settings dashboard.

All I did was added the folder to chrome extension (dev mode)

And the Ads Stopped

I was like what the heck in 2 mins and in just 1 step.

We are living in an Era where we don’t need to find the right software to download and setup anymore.

We can just build the software we need.

What a time to be alive.

[Raavi.dev @RSP_ER]

Hey @grok Conduct a serious, peer-reviewed psychological analysis of people who leaves 47 browser tabs open for 6 months "just in case they need them later."

Name this Mind Bug thing discovered and Coined by me.

[Raavi.dev @RSP_ER]

Indian Army just made Chinese Army Run away by just using fucking Sticks.

This Country is Here to Stay forever!!!

[Raavi.dev @RSP_ER]

Best personal agent right now depends on the job:

Professional work, docs, judgment & proactivity are better done by Claude Duet.

Coding, local files, computer use & voice are best done by dots with Codex.

Multi-step automation, webhooks & handoffs works great in Grokbot.

Everyday “just text it” tasks are great with Instinct or Muse.

[Riley Brown @rileybrown]

My unfiltered thoughts on personal agents.

Grokbot has the best UI and it’s arguably the best overall. In a lot of ways they’re ahead. Their desktop app feels the best, and their multi-bot coordination is great: you can set a team lead and it passes tasks to the other bots. Grokbot also has the best “webhook” feature, which means you can easily turn outside events into triggers that start your agents. That speeds up background workflows like watching your email, Slack and more. It also has Cursor integration, so for bigger tasks Grokbot can spin up a Cursor cloud session with any model (until OpenAI cuts SpaceX off from their models, which I think will happen soon). Unfortunately, running frontier models through Cursor costs more than Claude or Codex. Still, they’re the only ones embracing a team of agents. It’s a really cool idea, but the more bots you have, the more time you spend managing and maintaining them. I’m starting to think one bot with different threads/channels is the right answer, which I’ll get to with Claude Duet.

dots by OpenAI is rough around the edges right now, but it’s probably the most powerful if you use Codex, which I do. dots can spin up Codex sessions in the cloud or locally on your computer. So if you have coding or video editing workflows that need your local files, this is very useful. Codex computer use is also the best in the world: your dot can spin up a local Codex session that fully controls your computer, as long as it’s on. Their real-time voice is incredible. If you have your email and work tools connected, try this: go on a walk with headphones in, ask the agent to summarize your emails, have it draft replies and read them back to you, then send them off. If I remember right, this is close to a workflow you (Tim) used to have with your VAs before AI. It’s my favorite use case for personal agents. The one downside is that you have to call your agent, and it’s an awful experience. It literally rings, and it’s annoying.

Muse by Meta is a toy. I’ve talked to a bunch of finance people I play poker with who all tried it and set up automations, and they all said it doesn’t reliably do what they want. It’s fast, and its user numbers will probably keep growing. But it’s basically free: you get a TON of usage without paying a dime. That’s NOT a trade I’m willing to make. An agent doing high-leverage tasks should be running at the Opus 5.5 or GPT 6.1 Sol level.

Claude Duet - is in early access, so unfortunately not everyone can use it yet. First things first, it runs on Opus 5.5, the best model in the world. I’ve only had access for 72 hours, but Duet feels different from the others. It’s not a personified agent like Grokbot, dot or Muse, which gives it a more professional vibe that I like. When you ask your Duet agent to do something, it spins up a task, which is a new chat thread. All your threads show up in a Tasks tab, so it’s easy to see everything Duet is working on or has finished. And instead of creating more agents (like Grokbot), you create channels, just like Slack (I can tell @anthropic is Slack-pilled). I have a content channel, a finances channel, etc., which keeps different kinds of tasks separate. The only other app I’ve seen do this is Buzz by @jack, and it’s a good experience. Their strong point is documents (artifacts). If you want decks, high-quality docs or design mockups, Opus just wins. Period. It’s not close right now. Their weak point is real-time voice. It’s the worst of the bunch and borderline unusable, though I think they only added it a few days ago.

Instinct lives in iMessage, has no app, and doesn’t do anything notable besides show you product suggestions you didn’t ask for. (Not ads, according to their team.)

My hot take:
The VM is overrated. SpaceX, OpenAI and Meta all talk about the VM your agent can use and feature it in their marketing. It’s not as useful as you’d think, because more and more websites block bots that visit their site on the cloud VM. All of these companies promise their agents can make payments, but most payments either take forever or get blocked altogether. SpaceX has a cool workaround with Grokbot that routes traffic through your computer so it uses your IP address. But I’d much rather the agent just USE my computer for the task, like dot does by handing it off to Codex. Also, i don't think normal people care about the virtual computer. If anything they get offended they have to "stream a computer".

The problem with personal agents: by the time you read this, pick one, and test it for 7 days, every one of them will have released 10 new features and the whole landscape will have changed. It’s nearly impossible to keep up.

[Raavi.dev @RSP_ER]

This Special Article is about IOS App "BiteFix: Clean Grocery Scanner" https://t.co/Q6dCCcsiLO