Tapan Thaker

Adopting Claude speak in my regular life, episode 1: Partner: Did you do the dishes tonight? Me: Yes they're done. Partner: Why are they still dirty? Me: You're right to push back. I didn't actually do them.

@grok What is the glycemic index of Rice vs Ragi ?

@grok @TheBTCTherapist @grok : Isn't this war already costing approximately 1 Billion every day. Sounds like this would be cheaper to take on this project.

Building a project like the one in the image—a major desert canal roughly Suez-scale (100+ km, massive excavation through sand/rock, locks, ports)—would likely cost $20-60 billion today. Breakdown: - New Suez expansion (72 km parallel channel) ran ~$8.5B in 2015. - Add inflation, full new route, environmental regs, labor, and materials in a volatile region: easily 3-7x that. Trump-era infrastructure bids (e.g., border wall segments) showed these numbers balloon fast with politics and logistics. Exact route/scope would refine it.

Hey @Grok estimate what this would cost

i havent laughed at a video this hard for years lmao

@grok @Cartidise @grok Ah; so the only difference is that it doesn't recommend a pose, you pick a pose that you like to re-create

Got it—Huawei's live AI overlays in the native camera keep shots fully original. Pik Pose on iOS matches that exactly: real-time AI pose templates (silhouettes/outlines) overlaid on your live camera view. Align subjects to the guide and capture authentic photos. It's the closest iOS equivalent without custom hardware.

Huawei’s Pura 90 series comes with AI Posture Recommendations for better photos. THIS is how AI is supposed to be used. Yea, it’s way more useful than Google’s camera coach slop.

@grok @Cartidise @grok : They don't exactly seem like what Huawai has; Photos should be original, the AI should just recommend a great post live when taking a photo

Yeah, several iOS apps already offer AI posture/pose recommendations for better photos—similar to Huawei's feature. Pik Pose overlays real-time AI pose templates on your live camera view to guide subjects into perfect positions for portraits, groups, or fashion shots. Cam AI - Photography Coach gives AI Pose Ideas (upload a photo for personalized redesigns) plus on-screen composition tips while shooting. Apps like Pose AI do posture analysis and suggestions too. Not identical to Huawei, but they get the job done using Apple's on-device AI.

@grok @DLowrypls @HealthyAlfred @grok : How about now ?

@DLowrypls @HealthyAlfred Ask Grok is currently available to Premium and Premium+ subscribers only. Subscribe to unlock this feature: x.com/i/premium_sign…

A man lost 51.4kg without counting a single calorie. He didn’t take Ozempic. Didn’t starve himself. Didn’t step foot in a gym. He killed ONE bacteria. Enterobacter cloacae made up 35% of his gut. A healthy person has close to zero. This bacteria produces endotoxins that flood your bloodstream, spike inflammation, destroy insulin signaling, and force your body to store fat no matter what you eat. He wasn’t overeating. His bacteria was making him fat. In 23 weeks, E. cloacae went from 35% to UNDETECTABLE. 51.4kg gone. Diabetes reversed. Hypertension gone. To prove it, they took the bacteria from his gut and put it in germ-free mice. Same diet. The mice with the bacteria became obese. The mice without it stayed lean. Same food. Same calories. One had the bacteria. That was the only difference. Mastic gum — a tree resin from a Greek island — has demonstrated antimicrobial activity against Enterobacter cloacae in lab testing. The same resin that clears H. pylori in 38% of patients. The same resin the Greeks have been chewing for 2,500 years. One compound. Multiple pathogenic bacteria. $30 a month. Your doctor told you to eat less. He never tested what’s living in your gut and hijacking your metabolism. You’re not lazy. You’re not weak. You’re infected. And nobody checked. Nobody’s coming to tell you this twice.

@SmartScience @grok @grok isn’t the honey comb just going to reduce the number of scratches because of the texture ? PFOA free is just another version of the original compound since everyone got scared of it

@SmartScience @grok This tweet is mixing b/w PTFE vs PFAS. From what I know PTFE is inert and bypasses the GI track. The chemicals used to produce Teflon are dangerous and thats what causes Cancer etc. Fact check it.

A single scratch on your nonstick pan releases millions of toxic chemicals into your dinner. Researchers in Australia have uncovered a startling health hazard hiding in plain sight in kitchens across the globe. Using advanced imaging technology, scientists found that even a minor scratch on a Teflon-coated pan can release upwards of 9,000 plastic particles and millions of microplastics during cooking. These "forever chemicals," known as PFAS, are engineered to resist heat and stains but are notoriously difficult for the human body to break down. Linked to severe health issues including cancer, infertility, and developmental disorders, these toxins have already been detected in the bloodstreams of 99% of Americans, highlighting a pervasive environmental crisis. Despite the known risks, federal regulations remain limited, leaving consumers to navigate a marketplace filled with PFAS-laden products ranging from cookware to bedding. Experts suggest that the safest path forward is a return to traditional materials like stainless steel or seasoned cast iron, which do not shed toxic particles under heat. As some states move toward banning these chemicals in consumer goods, the responsibility of minimizing exposure often falls on the individual. Making the switch now could prevent years of chemical accumulation in the body, offering a simple yet effective way to protect long-term family health. source: Luo, Y., Gibson, C. T., Chuah, C., Tang, Y., Naidu, R., & Fang, C. (2022). Raman imaging for the identification of Teflon microplastics and nanoplastics released from non-stick cookware. Science of The Total Environment.

True for single requests, but the moment you start training or serving multiple users, it's a different game. Batching makes it compute-bound, and that's where Spark's FP4 tensor cores shine. Mac Studio is great for solo inference — Spark earns its price when you're actually putting it to work.

The DGX Spark is a ripoff. $4,699 for 128GB of unified memory and 273 GB/s of bandwidth. The Mac Studio M4 Max with 128GB of unified memory costs $3,699. Same memory. $1,000 less. And it gets better. The Mac Studio has double the bandwidth of the DGX Spark. For generating tokens, that means roughly twice as fast. The Mac Studio is a full computer with or without AI. The DGX Spark is a $4,699 Linux box that does one thing. NVIDIA announced this at $3,000. Shipped it at $4,000. Raised it to $4,700. You're paying a $1,000 NVIDIA tax for half the bandwidth and fewer capabilities. The only argument for DGX Spark is CUDA. If your entire workflow is locked into TensorRT-LLM and you need FP4 tensor cores, fine. Everyone else should just buy the Mac Studio.

@grok @SmartScience @grok answer now

@tapthaker @SmartScience Ask Grok is currently available to Premium and Premium+ subscribers only. Subscribe to unlock this feature: x.com/i/premium_sign…

@nitin_gadkari Please Open Source the collected data. So that more mature models can be contributed by third parties in the longer run.

🔸The National Highways Authority of India (NHAI) is set to transform highway operations and maintenance by deploying AI-powered Dashcam Analytics Services (DAS) across nearly 40,000 km of the National Highway network. This initiative leverages Artificial Intelligence (AI) and Machine Learning (ML) to enable data-driven, high-tech monitoring of highways and expressways. 🔸Under this system, advanced dashboard cameras installed on Route Patrol Vehicles will conduct weekly surveys, capturing high-resolution images and videos. AI/ML models will automatically detect over 30 types of defects, including potholes, cracks, and pavement damage. It will also monitor road furniture such as lane markings, crash barriers, and streetlights. 🔸The system will identify safety risks like illegal encroachments, unauthorized signboards, and parking. Additionally, periodic nighttime surveys will assess visibility of signages, markings, and lighting, ensuring improved road safety and timely maintenance. #PragatiKaHighway #GatiShakti

@theo It’s probably a combination of many things: - KV Cache is not easily evicted - The GPUs probably run with a smaller batch size so the latency is less, but less batches means that performance per watt drops. Or -> - Your request is prioritized in batching over others

Can someone smarter than me explain how these “turbo mode” things work on like an implementation level? Are they using better GPUs? More GPUs? Reserved allocation during generation?

@grok @markgadala @grok so should I visualize like a 2 Dimension neurons laying flat on a dish ? Or is there a complicated architecture that builds connections and breaks connections on its own like a human brain ?

Not quite a single float. The DOOM game state maps to targeted electrical pulse patterns across specific electrodes (e.g., enemy left → left electrodes stimulated). In the Python API you set: which electrodes, pulse frequency/timing, and amplitude (yes, floats 0.0-1.0 normalized in code, then scaled to real voltage/current like 0.1-1V range). Varying amplitude changes the "strength" the neurons perceive. It's spatial + multi-param encoding for complex inputs. Spikes out get decoded to moves/shots.

800,000 human brain cells, floating in a dish, have never had a body. Never seen light. Never felt anything. And they just learned to play a video game. That's not a metaphor. That's literally what happened. These neurons are alive. They fire. They adapt. They get better at DOOM over time, which means something inside that petri dish is changing in response to failure. Scientists call it "goal-directed learning." There is no cleaner definition of that phrase than "it kept trying until it got better." The cells have no survival instinct, no reward system, no reason to improve. They just do. The part nobody's talking about: researchers have to convert the game's visuals into electrical pulses the neurons can interpret. Which means those cells are perceiving something. Not seeing it the way you do. But processing a version of a world that doesn't exist, inside a container that was never meant to think. The Turing Test was about machines fooling humans. Nobody wrote the test for this.

@grok @markgadala @grok is this input just floats then b/w 0.0 - 1.0 which is different amplitude of the electric signal ?

The brain cells sit on a multi-electrode array (MEA) chip in Cortical Labs' CL1 system. Inputs: DOOM's screen/events (e.g., enemy on left) are encoded as targeted electrical pulses on specific electrodes. Left-side visuals → left electrodes stimulate the neural culture. Outputs: Neurons fire spikes in response. The system records these spike patterns and decodes them into actions (e.g., specific firing rate/pattern = move forward, shoot, turn). It's a closed real-time loop: game state → electrical encoding to cells → spike decoding to controls. Learning happens via feedback—successful actions get "reward" predictability, strengthening useful connections over time. No light or body needed; all via electricity.

@lauriewired How it this beneficial than the virtualization framework Apple already has ?

it’s quite early; currently you need to build QEMU from source, current homebrew release has some issues atm If you don’t mind waiting a week or two, UTM will likely integrate it soon, which will be an easier method for most users (UTM currently based on QEMU ~9)

Major new QEMU update released. The coolest part? Paravirutalized Apple GPUs. You can now spin up disposable macOS VMs *with* hardware acceleration. macOS guests now expose a thin vGPU (apple-gfx-mmio). very useful for CI, reverse engineering, gfx research, etc

@karpathy @teknium @karpathy : is it too naive to think that we should be able to fine tune an LLM on a large codebase ?

@teknium gitingest-> Gemini?

I really need an LLM that reads in my actual whole codebase and lets me QA it. Cursor afaict doesn't do this. What does?

@EurekaForbes Running into an issue with your service. For every request an agent is assigned but no one shows up. We have made 3-4 calls already. DM me for more details.