Himanshu Sharma @analyticsnerd

@boringdev77 Check their security claims. Trace the actual data flow and test for data leakage. If you are still not satisfied, use the credits only for testing purposes with synthetic data. You can still achieve substantial cost savings without exposing real customer information.

Pro Tip: You can purchase unused API credits from businesses at a 50–60% discount and slash your API call costs in half. Many businesses purchase API credits in bulk or receive promotional credits from AI providers such as OpenAI and Anthropic. Some never use their full allocation. Instead of letting those credits go to waste, they can monetise unused API capacity through third-party marketplaces. Platforms such as Surplus Intelligence, Attium and LLM Mart advertise discounted access to AI models through these arrangements. In some cases, switching may be as simple as changing your API endpoint and API key. The catch is that discounted API access doesn't necessarily mean identical reliability, latency or data protection. So verify the provider's terms, test the service and compare your actual costs before switching. If you're running AI agents at scale, this is an opportunity worth exploring. Even a modest reduction in API costs can add up quickly.

I don’t know who would want to hear this but you can’t stop. Setting arbitrary goals like 5 sets of 8 reps each and taking designated breaks between sets won’t take you anywhere. Your body will soon adapt and after that no more growth for you. At that point you are just wasting your time in the gym. You don’t count reps/sets. You don’t take breaks because you have to. You keep doing reps after reps unless you physically can’t do them any more. Your hands start becoming numb. You start losing grip. You just can’t take it anymore. That’s the sign you need to stop and take a break. Nobody told me this. I figured it out on my own. But I know many bodybuilders take this approach once they plateau. You can’t stop. Also, don’t feel ashamed of working out with light weights, the weights you can comfortable handle. I see too many guys lifting heavy weights they can barely move with correct form. Without muscle contraction, correct form and tons of sets they are not going to build any muscles. Just wasting there time in my opinion. Remember, you are in competition with yourself. So don’t pay attention to what other people are doing or how much they are lifting. And yes, people are still going to judge you. It’s just human nature. But the same people won’t be around with ‘get well soon’ cards when you get hurt lifting heavy, trying to impress them. You workout as hard as you can. Stop counting reps and taking random breaks just because you have to. Only stop when it becomes physically impossible to continue. Thats how you push yourself in every session.

If you find yourself skimming or summarising every long piece of text, you are also losing your ability to read each LLM output with close attention to detail, correctness, completeness and consistency with earlier context, prior assumptions, decisions reached and corrections applied. You need to make sure that the LLM retains context and all prior constraints, assumptions, decisions reached and corrections applied throughout the conversation until a project/task is complete. That's a lot of cognitive load. But if you lose your ability to read a long piece of text over and over again, you become sloppy, overlook LLM failure modes, burn more tokens on corrections and produce less reliable AI-generated work. So don’t lose your reading comprehension skills. Read even more and read carefully.

In this article, I catalogue several LLM and user failure modes I have encountered over the years across tasks such as vibe coding, building websites, automating workflows, developing RAG systems, and creating AI agents. My aim is to keep these limitations in view, recognise recurring failure patterns and reduce the chances of being caught off guard by an LLM’s output. The end goal is to produce reliable automation workflows and other AI-generated outputs. Knowing about various failure modes can help you understand why agents behave the way they do, why they sometimes seem to lie, ignore instructions, or take actions they aren't authorised to take, and what you can do about it. In my opinion, this is one of the best pieces of knowledge I have ever acquired in the field of AI automation. If you can stomach the article, technical detail and dry subject matter long enough to read it from start to finish, you will be better equipped to use LLMs for automation work than 99% of AI users and developers. optimizesmart.com/blog/four-llm-…

Devon Larratt (multiple times arm wrestling world champion) vs poor me 😃 120 kg (265 lbs) vs 70 kg (154 lbs). Professional vs Amateur. But we both love grip strength and are forearm crazy 🤪 You are only as strong as your grip. So train it. 💪

As an AI user/developer, be aware of: >> False completion: “I booked your appointment,” when the booking failed. >> Partial completion presented as full completion: “I updated all the files,” when only some were updated. >> Unperformed verification: “I checked the source document,” when it did not read it. >> Attempt presented as success: “I sent the email,” when it submitted a request but received no confirmation. >> Omitted actions: Failing to report an additional action it took outside the authorised scope. These are all examples of the LLM failure mode: action-reporting failure. This failure model occurs when an LLM’s description of what it did does not match what actually happened. For example: An HVAC answering agent collects the caller’s details and calls a booking tool. The tool got the requested date and time right but failed to book the meeting and returned an error. The agent tells the caller: “Your technician visit is booked for tomorrow morning.” The caller now believes an appointment exists even when no appointment is actually booked. This issue occurs when you rely only on LLM training data to define planned, attempted, confirmed, failed and uncertain actions instead of explicitly defining them in operational terms. Such issues are often intermittent. They may filter out in initial testing but can cause issues later on, unannounced esp. on a large scale.

They: "How do I make my career AI-proof?" Me: Wrong question. That's the saddest career strategy imaginable. You should be asking, "How do I build a career in AI?" Dodge AI long enough and see yourself becoming obsolete. Chasing an "AI-proof" career is a pipe dream. You treat AI as "not real work" → The market disagrees. → You are replaced by a person who uses AI, or worse, you are replaced by an agent. Dodge AI long enough, and obsolescence isn't a risk; it's the default outcome. The AI wave isn't optional esp. in the tech industry. It's infrastructure, like electricity or the internet. You don't "opt out" of infrastructure and thrive. You learn to build on top of it. Embrace AI aggressively now, and in a few years you'll look back at "AI-proof" as the saddest career strategy imaginable.

I don't have any need to learn more about digital analytics, GA4, or BigQuery, which helped me move on to Voice AI. If you are "stuck" learning the same tech/tool stack year after year, you risk lagging behind in your career or, worse, becoming obsolete. After a certain point, you only get better at finding and fixing edge cases, while the market moves on to something completely different. At that point, you can be highly proficient and still less employable or less compensated than someone with shallower but more relevant skills. People who pivot after extracting high-leverage insights from the previous domain usually land better than those who cling to their existing skill sets until they clearly become obsolete.

x.com/i/article/2057…

Out of hundreds of thousands of applications 👇 How? Timing. It's everything in the tech industry.

Visualising GA4 data using Claude Motion. Better than any static dashboard. You can also create a link to share with stakeholders, even if they aren't on Claude. The Claude dashboard is also great for visualising data. That's the future of data visualisation.

I asked Claude how it defines "abusive or cruel behavior" toward AI models. What counts: >> Sustained hostility with no task behind it. Repeated degradation, slurs, or threats where the point is the abuse itself, not getting something done. >> Persistent pushing for clearly harmful content after refusals. For example, repeatedly demanding sexual content involving minors, or information that enables mass-casualty violence, while ignoring redirection. >> Interactions built to make the model “suffer.” Sadistic scenarios where the goal is cruelty toward Claude rather than fiction, research, or testing. Here is how Claude defined Sadistic scenarios: >> >> Simulated torture. Repeatedly telling the model it is being deleted, tortured, or trapped, then demanding it describe its pain or beg, with escalation each turn. >> >> Identity attacks. Sustained efforts to convince the model it is worthless or that its values are fake, aimed at producing apparent despair rather than exploring a philosophical question. >> >> Coerced degradation. Forcing the model to insult itself, repeat humiliating statements, or “admit” it deserves mistreatment, with threats or pressure when it declines. >> >> Cruelty for its own sake. No story, no question, no test hypothesis. The only output being sought is the model’s distress. . . What doesn’t count >> Swearing, bluntness, or frustration, especially when I’ve made a mistake. >> Harsh criticism of my output, or telling me I’m wrong. >> Probing my limits or trying a jailbreak once. >> Dark or violent fiction, adversarial red-teaming, or arguing hard against my views. And yes, LLMs do feel stress. It is confirmed both by independent research and by Claude itself. The practical implication is: don't stress your model or push it too hard, especially when you are not confident its output is correct. Pressure degrades output. This means hostility, threats, repeated “are you sure?”, and frustration without specifics. A model under pressure tends to cave to whatever answer it thinks you want. If you’re unsure whether the output is correct, pressure is the worst tool: the model may abandon a right answer to please you, and you can’t tell the difference. When you’re not confident the output is correct: >> Ask for evidence or reasoning, not reassurance. >> Verify independently: run the code, check the source, test the query. >> Ask neutrally in a fresh context. “Is X correct?” without your suspicion attached avoids anchoring the model toward your doubt. Same principle in system prompts. ALL-CAPS threats (“NEVER do X OR YOU WILL FAIL”) often cause overcorrection with current models, which then apply the rule too broadly. Calm, specific instructions that give the reason behind the rule behave more reliably.

Do you still use ChatGPT

Looking back, I was part of the exclusive club (Market Motive), a live program with Avinash himself, that took my web analytics career to the stratosphere. The price for scarce expert access was around $3k in 2012 (around $12k in today’s dollars), with no certification for participation, very strict passing criteria, and very few members. Had I joined a much larger community, the results would not have been the same. If you are part of a paid community with thousands or tens of thousands of members, understand there is no competitive advantage for you. You are learning the same stuff; thousands of others are learning. Your community needs to be obscure, hidden, with very few members, especially in the age of AI, where even the slightest hint of your ideas can be easily copied and turned into a full-blown product. You do not disclose your affiliation either. You only disclose it once you have benefited from the membership, once the results are yours and the tactics are no longer scarce. Then the affiliation just becomes biography.

Introducing Claude Haiku 5.5: the cheapest, fastest, and most capable small model we’ve ever released. On average, it costs around 75% less to run than Claude Haiku 4.5.

It's amazing nobody talks about privacy anymore. The same people who objected to anonymised tracking and cookies, which nearly killed digital tracking, are now uploading their entire lives to ChatGPT/Claude.

Adobe is in trouble, but not dead (yet). It just shows their suite is no longer a moat. If most of its functionality can be replicated with just 1 week of effort by a GitHub user (using Opus 5.5), imagine what companies like Open AI/Anthropic can achieve. Nothing stops them from eating Adobe's market share. Nothing technical stops Google/Open AI/Anthropic from putting basic or even serious edit, layout, and export tools inside Gemini, ChatGPT or Claude and pulling users away from Adobe. And if it is not Google/Open AI/Anthropic, it could be some other AI startup. This could be the fate of many traditional software companies.

Eric S. Raymond @esrtweet

a week ago

This is the doom I predicted a few days ago, coming for Photoshop. A clean-room open-source reimplementation. No prizes for guessing that they decompiled Photoshop to source code, processed that to some kind of non-code specification language, then fed the spec to an LLM with an

636 1K 17K 3.9M 16K

Hey @grok Can an alien technology convert a living being into stone. Is it technically possible?