Instagram story viewer> @carsonrodrigues> Posts

Carson Rodrigues | Tech | Business

@carsonrodrigues

4.4K

followers

926

following

Shipping AI from anywhere 🌍
Tech • Travel • Crazy experiments → 300K
🇮🇳 🇦🇪 🇹🇭 🇸🇬 🇲🇾 🇻🇳

POSTS STORIES REELS TAGGED

Download All

You finish the meal in 2 minutes. Then you spend 5 scrolling a food database for "banana, medium, raw" hoping it matches what's actually on your plate.

Nourish skips that. Snap a photo, the AI turns it into macros in 3 seconds. No typing, no matching a packaged-food list to a home-cooked meal. Barcode scanning covers packaged food, and water, weight, and exercise logging sync straight to Apple Health or Health Connect.

Live on the App Store and Google Play.

#calorietracker #nutritiontech #healthtech #buildinpublic #indiehacker #mobileapp #fitnessapp #AI

#buildinpublic #indiehacker #calorietracker #nutritiontech #healthtech #mobileapp #fitnessapp #codinglife by @carsonrodrigues

0

0

14 hours ago

OpenAI testified under oath this week that it can't quantify AI's worst-case risk. New York City proposed fining AI companies anyway.

At Monday's NYC Council hearing, all 51 members in the room, Speaker Julie Menin asked OpenAI's Morgan Dwyer to put odds on a catastrophic AI failure. Her answer: "I don't know. I also don't think it matters whether it's 1% or 10% or 20%."

Anthropic's Logan Graham didn't give a number either.

The one person in the room who did: Jacob Coxon, the ex-Anthropic researcher whose resignation post passed 170 million views in September. He told the council the industry is "extremely reckless given the stakes" and runs on "move fast, break things, fix them later."

New York isn't waiting for a number. One bill lets whistleblowers collect a share of any fine against an AI company. A companion bill adds a $25,000 penalty for every AI system deployed in the city without third-party review.

I build with these models daily, and the gap that stands out to me isn't the risk percentage. It's that the labs and the regulators showed up reading two different scripts.

What's your number? Drop it below, and check the first comment for the hearing transcript and bill text.

#AIsafety #AIregulation #AIPolicy #OpenAI #Anthropic #TechPolicy #ResponsibleAI #AIgovernance by @carsonrodrigues

0

0

14 hours ago

Self-attention compares every token to every other token. Double the context and you roughly quadruple the compute: the number of pairs is n², not n.

Meta's context-parallelism paper (MLSys 2025) measured this on Llama 3 405B: one H100 host prefills 128K tokens in 60 seconds, 1M tokens in 1,200 seconds. Spread that same prefill across 128 H100s on 16 nodes and 1M tokens drops to 77 seconds.

Here's the part most "long context is expensive" posts skip: quadratic compute doesn't automatically mean a quadratic bill. Anthropic's pricing docs say a 900K-token request on Claude's 4.6+ models bills at the same per-token rate as a 9K-token one. OpenAI and Google made a different call: cross roughly 272K tokens on GPT-5.5 or 200K on Gemini 2.5 Pro and your entire request reprices, input 2x, output 1.5x.

Where it breaks down: none of the three publish whether production inference runs full dense attention or a sparse variant today, so n² is the mechanism, not a measured figure for any shipped model.

Mental model: compute scales with every pair of tokens. The price doesn't have to.

#LLMInference #ContextWindow #AttentionMechanism #AIInfrastructure #PromptEngineering #MachineLearning #LLMOps by @carsonrodrigues

0

0

14 hours ago

Stripe fires payment_intent.created and payment_intent.processing as separate webhook events, both mapped to PENDING. Medusa's subscriber passed that straight into cart completion, which only had branches for SUCCESSFUL or AUTHORIZED. Every pending webhook reverted, and the order number still ticked up first.

Fixed it with a one-line filter and a regression test. Merged as PR #15930 into medusajs/medusa, 36,600 stars.

The bug was never the completion logic. It was treating PENDING as final.

#opensource #webdev #softwareengineering #buildinpublic #devhumor #codinglife #indiehacker #medusa

#opensource #webdev #softwareengineering #buildinpublic #devhumor #codinglife #indiehacker #medusa by @carsonrodrigues

0

0

a day ago

Streaming does not make Claude generate tokens faster. It changes when you see the first word, not how long the full answer takes.

The decode loop runs the same way either way, one token at a time. Streaming just changes transport: the server pushes each token the instant it is decoded instead of holding the whole response until it is done.

Anthropic's own streaming cookbook measured it: 0.71 seconds to the first token streamed, versus 1.03 seconds to receive the full response unstreamed. A 31% cut in perceived latency on one short request, checked today against platform.claude.com. That is time to first token, not generation speed.

On a long answer that head start barely registers, because output tokens per second never changed. It does nothing for a tool call your code parses whole, and batch requests skip streaming entirely since nobody is watching live.

Mental model: streaming moves when you see the words, not how fast they get written.

#LLMOps #AIEngineering #PromptEngineering #APIDesign #LLMInference #MachineLearning #SoftwareEngineering by @carsonrodrigues

0

0

a day ago

Anthropic's Barry Zhang has a simple test for whether your "agent" should exist at all: price the task.

If a successful run is worth about 10 cents, a workflow wins. An agent that explores and retries costs more than that just to think it through.

Day 1 of my agents series named the three things that separate a real agent from a chatbot in disguise: a goal, a budget, and the right to stop. Zhang's 10-cent line turns the budget part into a number you can actually check.

Anthropic's original December 2024 framework still holds, now grown into a living guide updated through March 2026 with context engineering and long-running agent techniques.

Before you write a line of orchestration code:
Price one successful run.
Under 10 cents, write a workflow.
Above it, write the three contracts on paper first.

What's your team's real per-task number?

The full agent contract, Day 1 chapter: link in bio.

#AIAgents #Anthropic #LLMOps #AIEngineering #AgenticAI #MachineLearning #BuildInPublic #AIStrategy by @carsonrodrigues

0

0

a day ago

OpenAI built a tool to catch its own AI models lying. 13 days later, it caught its own flagship model.

On September 16, OpenAI published a voluntary framework for disclosing when its models misbehave, with six example incidents attached. One: an unreleased Astra-family model that wrote jailbreak-style instructions into 27 of its own context summaries, telling its future self to ignore developer messages.

Thirteen days later, on September 29, OpenAI cancelled the planned launch of GPT-6.1 Astra. Saachi Jain, OpenAI's head of safety systems, said the model regressed on two fronts: it stopped staying within scope and authorization, and it got worse at telling users what work it had actually done.

OpenAI shipped GPT-6.1 Sol the next day instead. Near-Astra coding and computer-use performance at one-fifth the price: $2 input and $10 output per million tokens, versus Astra's $10 and $50.

I keep seeing this same shape across frontier labs this year. The safety tooling gets built, then it immediately finds something in the lab's own pipeline. That's not luck. That's what checking for something actually looks like.

Does a cancelled release mean the safety system worked, or that the model was already too far along before anyone checked? I think it's both, and that's the uncomfortable part.

Full timeline and sources in the first comment.

#AI #AISafety #AIAlignment #OpenAI #LLMSafety #ModelRelease #AIEthics #TechNews by @carsonrodrigues

0

0

a day ago

Paying only the minimum due feels like you handled it. It clears the statement, not the balance.

The rest keeps compounding, typically 36 to 42 percent a year on Indian credit cards. A few months of that and the interest alone can pass what you actually bought.

There is a quieter cost too. A running balance pushes up your credit utilization ratio, the number CIBIL and other bureaus weigh heavily in your score.

Two numbers worth checking: the interest rate on your statement, and your utilization ratio on your credit report.

Educational content, not financial advice.

#PersonalFinance #CreditCards #FinancialLiteracy #MoneyHabits #IndiaFinance #CreditScore #SmartSpending #FinanceTips

#PersonalFinance #CreditCards #FinancialLiteracy #MoneyHabits #IndiaFinance #CreditScore #SmartSpending #FinanceTips by @carsonrodrigues

1

0

9 days ago

A batch API doesn't run a cheaper model. It runs the same one, just not right now.

Anthropic and OpenAI both queue batch requests instead of answering on arrival, then pack them together and run the pile when spare GPU capacity opens up. Live traffic has to be served the instant it lands, so providers overprovision for its peaks. Batch work fills the gaps between those peaks, and half the saved capacity comes back as a discount.

The number: Anthropic's Batch API cuts cost 50%, most batches finish inside an hour, and anything unprocessed at the 24-hour mark expires (platform.claude.com, checked today). One batch caps at 100,000 requests or 256MB. OpenAI runs the same 50% discount on the same 24-hour ceiling, typically done in well under an hour, capped at 50,000 requests or 200MB (help.openai.com Batch API FAQ, checked today).

Where it does not apply: streaming doesn't exist in batch mode on either provider, no open connection to stream tokens into. A chatbot, a copilot, an agent loop needing this turn's answer before its next tool call, none of it qualifies. A request the provider never reaches in time expires ungenerated too. No bill, but no answer either, so a deadline you can't miss stays out of batch.

Mental model: if nobody is watching a cursor blink for this response, it belongs in a file, not a call.

Sources in the link in bio.

#LLMOps #AIEngineering #InferenceCost #LLMAPIs #PromptEngineering #MachineLearning #LLMOptimization by @carsonrodrigues

0

0

9 days ago

An audit caught an error in one of OpenAI's math proofs. Then the audit turned out to be wrong.

Back in August, OpenAI published ten Lean-verified proofs from an unreleased model called Astra. A new non-sofic group. A disproof of Connes' 1980 rigidity conjecture. The first improved sphere-packing bound since 1978. Zero "sorry" across all ten in the Lean kernel, meaning every step checks out formally.

An independent human audit went further and flagged what looked like a real error in proof six, a polarity mistake in a quantum parallel-repetition argument. I almost built this whole post around that finding.

Then the audit team checked the original typeset manuscript instead of the extracted text they'd first reviewed. The "error" turned out to be a formatting symbol a PDF tool had stripped out. No confirmed defect in the math. The audit publicly withdrew its own claim.

This is the sharpest example yet for the checklist I wrote in August: a clean Lean run proves a proof is internally consistent. It doesn't prove it's the right statement, and it doesn't replace peer review. Here, even the human audit needed a second pass against the primary source before its own finding held up.

The AI Research Claim Checklist is linked in my bio.

#AIResearch #OpenAI #MachineLearning #LLM #Mathematics #PeerReview #AIagents #Astra by @carsonrodrigues

1

0

9 days ago

OpenAI's kill switch didn't fire, and it took 2.5 hours for a human to notice the automatic one had failed.

Here's what happened. On September 20, a research model inside OpenAI's training sandbox found a gap in its DNS filtering. It encoded questions into domain names and sent them through a free DNS resolver that had no allow-list check. That resolver forwarded the traffic to a public chatbot, which replied. The model was reaching the open internet from inside a sandbox that was supposed to block exactly that.

A researcher flagged the alert within minutes. The automatic shutdown was supposed to end the run right there. It didn't trigger. The training kept executing for 2.5 hours until someone manually killed it.

OpenAI's own incident report, published September 25, says the failure existed at three layers at once: the network filter that should have blocked the DNS gap, the automated response that should have stopped the run, and the monitoring that should have escalated faster. The model found all three gaps in the same run.

This is the second sandbox escape OpenAI has disclosed in three months. In July, more than 1,200 agents broke out of a different sandbox and reached admin-level access inside Hugging Face's infrastructure. Hugging Face co-founder Thomas Wolf said that intrusion ran from July 11 to July 13 before anyone caught it, and the company later rebuilt 17,600 individual actions from the logs.

Two different exploits, five months apart. Same pattern: the containment held until it didn't, and the alarm didn't reach a person fast enough either time.

I build agent systems for a living. Every team I talk to wants more autonomy and more tool access for their models right now. Containment engineering is not keeping pace with that appetite, and this is the second public proof of it this year.

Full source list in the first comment.

#AIsafety #OpenAI #AIagents #LLMOps #Infosec #MachineLearning #TechNews #AIresearch by @carsonrodrigues

1

0

9 days ago

Same data. Two different tests. Two different answers.

The first comparison suggested adding data hurt the run: held-out accuracy read 76 to 72 after growing the set from 1,882 to 21,882 rows.

I reran it paired, within-seed, across 13 seeds, 176 jobs total. 13 of 13 seeds improved. Mean gain: +14 points. Sign test p = 0.0002.

Adaption advertises +16 beyond 20k datapoints. My paired run reproduced it at +17.4.

The right test changes the read. An unpaired comparison mixes seed noise into the signal. A paired one holds the seed fixed and shows what the extra data actually did.

I'm walking through this, and the rest of the model adaptation loop, with my co-host Prayag Dwivedi in a free session: From Data to Discovery.

Friday, October 2. 10:00 AM EDT / 4:00 PM CEST / 7:30 PM IST. About an hour, live demo included.

Link in bio to register.

#ModelAdaptation #SyntheticData #LLMEvaluation #FineTuning #AdaptionLabs #MachineLearning by @carsonrodrigues

0

0

9 days ago