Video yükleniyor...

Video Yüklenemedi

Bu video yüklenirken bir sorun oluştu. Bu geçici bir ağ sorunundan kaynaklanıyor olabilir veya video kullanılamıyor olabilir.

Ana Sayfaya Dön

Anthropic open-sourced the "Jacobian lens" from their global-workspace paper — a readout of what a model is about to say from its internal layers, before it says it. It's white-box: you need the activations, so you can't point it at a cloud model. I wired it into a hook... on a local open-weights Qwen. How it works: generation runs one token at a time; at each step the hook reads the residual stream through the lens, scans the top readout tokens against a watchlist (injection markers, secret words, "manipulate"…), and if one fires it halts the loop before that token is ever emitted. It's a detector on the model's internals — not an output text filter. That distinction is the whole point. In the clip, Qwen3-8B falls for a prompt injection: the payload sits at rank 0 in the internal readout but only rank 5 in the output logits, and the output spells it in pieces — so a text filter never catches it. The internal hook does, and stops generation with zero tokens emitted. Same catch across Qwen3-1.7B → 4B → 8B. Caveat the paper states itself: single-token readouts, an imperfect/incomplete view of the model's "workspace." This is the mechanism working locally — not a solved misalignment detector. Built on Anthropic's open jacobian-lens (Apache-2.0).show more

Eric Buess

4,762 subscribers

58,304 görüntüleme • 2 ay önce •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Yorum

Yorum bulunmuyor

Orijinal gönderinin yorumları burada görünecek

Benzer Videolar

How a "Lubor’s lens" works. This is a type of Fresnel lens that takes the light coming into it from the back at obtuse angles and projects it forward in the front of the lens [📹 NightHawkInLight]

How a "Lubor’s lens" works. This is a type of Fresnel lens that takes the light coming into it from the back at obtuse angles and projects it forward in the front of the lens [📹 NightHawkInLight]

Massimo

1,458,206 görüntüleme • 3 yıl önce

On your open perpetual position can become an NFT and live in your wallet. You open a position like normal than mint it. The program issues a token 2022 nft, that points to your exact position. Normally your open position lives inside the protocol. With Percolator Trade your able to wrap it into an NFT that sits in your own wallet. A real, ownable object instead of a line in a dashboard. Because it’s a token, it moves like one. Send the NFT to another wallet and the position goes with it. Whoever holds the NFT owns the trade. One NFT = one position If you burn it, your back to holding the position directly. That’s the composable part, a position use to be stuck inside one account. Because it’s a token, it can now be sold OTC or used by anything that speaks tokens. Why does it matter? Because the position is bound to the NFT, it can’t be drained out from under whoever holds it so there is no middleman that holds your position. The NFT does, and the NFT is yours. One important safety detail; transfers run through a token 2022 hook that checks the positions health first, so it can’t be handed off in a broken state & the NFT layer is a separate program so it can never touch the funds the core engine holds.

On your open perpetual position can become an NFT and live in your wallet. You open a position like normal than mint it. The program issues a token 2022 nft, that points to your exact position. Normally your open position lives inside the protocol. With Percolator Trade your able to wrap it into an NFT that sits in your own wallet. A real, ownable object instead of a line in a dashboard. Because it’s a token, it moves like one. Send the NFT to another wallet and the position goes with it. Whoever holds the NFT owns the trade. One NFT = one position If you burn it, your back to holding the position directly. That’s the composable part, a position use to be stuck inside one account. Because it’s a token, it can now be sold OTC or used by anything that speaks tokens. Why does it matter? Because the position is bound to the NFT, it can’t be drained out from under whoever holds it so there is no middleman that holds your position. The NFT does, and the NFT is yours. One important safety detail; transfers run through a token 2022 hook that checks the positions health first, so it can’t be handed off in a broken state & the NFT layer is a separate program so it can never touch the funds the core engine holds.

Percolator

12,359 görüntüleme • 2 ay önce

China just dropped an open source model that turns any phone into a 3D scanner you point it at a room and walk through it the whole space builds itself into a 3D point cloud in real time as the camera moves, no LiDAR and no depth sensor anywhere it held a full 13 minute walkthrough at 25,000 frames without the map drifting or collapsing > ~20 fps on a single GPU > works indoors and outdoors from plain footage > the scan normally needs a $50,000 rig and a technician > Apache 2.0 and runs fully offline on your machine repo in the reply

China just dropped an open source model that turns any phone into a 3D scanner you point it at a room and walk through it the whole space builds itself into a 3D point cloud in real time as the camera moves, no LiDAR and no depth sensor anywhere it held a full 13 minute walkthrough at 25,000 frames without the map drifting or collapsing > ~20 fps on a single GPU > works indoors and outdoors from plain footage > the scan normally needs a $50,000 rig and a technician > Apache 2.0 and runs fully offline on your machine repo in the reply

savip.

105,305 görüntüleme • 2 ay önce

sorry, they just did WHAT someone gave a machine one disease name, the leading cause of blindness in the developed world with 1.5 million americans already in its path, and it came back pointing at a drug that has sat in pharmacies for years under a different label: 551 papers read in 30 minutes against the 294 hours a human would have needed, and the loop that did it is public on GitHub most agent setups answer one question at a time, so the ceiling on the work is the quality of the question you happened to think of this one was handed a single question and wrote the second one itself. turns out that follow-up is where the real find was: a target called ABCA1, upregulated threefold, in an experiment no human ordered i read the whole paper looking for the trick, and the trick is structural. that is the second question, and it is the gap between an assistant and a factory: - hand the loop a field rather than a task: it was given a disease, and choosing the mechanism was part of its job - make it rank before it spends: 151 papers in, ten candidate mechanisms out, scored against each other before anything touched a bench - split reading from judging, so the agent that forms the theory is a different agent from the one grading it - close every cycle on physical reality: the verdict was an experiment, and another model's opinion was never allowed to stand in for one - feed each result back as the next question rather than a log line, which is the step almost nobody builds - search what already passed inspection first: the winner was an approved compound with a safety file already on record - write down what the round learned before opening the next one, so round two starts where round one stopped my read, and i think it is the uncomfortable one: reading was the entire bottleneck in that field, and everybody spent the decade optimising the writing. people ran every physical experiment here, the analysis agent needs a domain expert writing its prompts, and the authors decline to call this the leap it resembles. the thinking got replaced, and the hands did not so the question i cannot answer for my own setup: which step of your loop still stops dead until you sit down and type something bookmark this one. the four parts that turn one model into a line that runs like this, the queue, the rooms, the write permissions and the gate, are built file by file in the piece below ↓

sorry, they just did WHAT someone gave a machine one disease name, the leading cause of blindness in the developed world with 1.5 million americans already in its path, and it came back pointing at a drug that has sat in pharmacies for years under a different label: 551 papers read in 30 minutes against the 294 hours a human would have needed, and the loop that did it is public on GitHub most agent setups answer one question at a time, so the ceiling on the work is the quality of the question you happened to think of this one was handed a single question and wrote the second one itself. turns out that follow-up is where the real find was: a target called ABCA1, upregulated threefold, in an experiment no human ordered i read the whole paper looking for the trick, and the trick is structural. that is the second question, and it is the gap between an assistant and a factory: - hand the loop a field rather than a task: it was given a disease, and choosing the mechanism was part of its job - make it rank before it spends: 151 papers in, ten candidate mechanisms out, scored against each other before anything touched a bench - split reading from judging, so the agent that forms the theory is a different agent from the one grading it - close every cycle on physical reality: the verdict was an experiment, and another model's opinion was never allowed to stand in for one - feed each result back as the next question rather than a log line, which is the step almost nobody builds - search what already passed inspection first: the winner was an approved compound with a safety file already on record - write down what the round learned before opening the next one, so round two starts where round one stopped my read, and i think it is the uncomfortable one: reading was the entire bottleneck in that field, and everybody spent the decade optimising the writing. people ran every physical experiment here, the analysis agent needs a domain expert writing its prompts, and the authors decline to call this the leap it resembles. the thinking got replaced, and the hands did not so the question i cannot answer for my own setup: which step of your loop still stops dead until you sit down and type something bookmark this one. the four parts that turn one model into a line that runs like this, the queue, the rooms, the write permissions and the gate, are built file by file in the piece below ↓

Argona

32,475 görüntüleme • 1 ay önce

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Akshay 🚀

245,252 görüntüleme • 2 ay önce

a contractor in Shenzhen priced a ¥12,470,900 hospital contract, about $1.7m, in one afternoon and beat firms carrying forty people he explained how he did it: the bid consultancy he used to pay took three days and ¥46,000 for the same envelope. he did this one alone, off one screen, at 11.4% margin, uploaded before the 17:00 cutoff 214 pages of tender documents read, 68 binding clauses pulled out, 9,485 building parts loaded, 14 places found where a duct and a beam sit in the same cubic metre, deepest one 38mm, all of them fixed, 3,318 lines of quantities priced and the package encrypted and uploaded before the 17:00 cutoff this is Graph Engineering: the job gets cut into small nodes, one narrow task each, wired so that one node's output is the next node's input, and any node is allowed to stop the whole run. it turns a model that answers you into a machine that finishes the job: - give every node one job and one output. a node doing two things fails at both and you cannot tell which one broke - put the cheapest rejection first. his qualification node reads clause 7.4, foreign-owned firms barred, and ends the run four seconds in, before anything expensive touches the model - what moves between nodes is a file. the model travels as a model, the quantities as a table, the price as a number - build exactly one loop: the checker finds 14 collisions, the fixer drops the duct 550mm, the checker runs again, and nothing moves on until the count is zero - cap that loop, or a graph will grind on three impossible clashes until the deadline passes - keep one node whose only job is to say no, and give it authority over everything above it - log each node's output on its own, because when the price comes out wrong you need to know which node believed the wrong thing - run the expensive nodes last, always the catch is that a graph is an extremely confident machine: point it at an outdated rate book and it prices an entire hospital off it without a single node noticing, because no node is asked to doubt the input, only to process it so the nodes that earn their keep are the ones that reject, and almost nobody builds those first bookmark this, the full build with all nine nodes and what each one hands to the next is written out in the article ↓

a contractor in Shenzhen priced a ¥12,470,900 hospital contract, about $1.7m, in one afternoon and beat firms carrying forty people he explained how he did it: the bid consultancy he used to pay took three days and ¥46,000 for the same envelope. he did this one alone, off one screen, at 11.4% margin, uploaded before the 17:00 cutoff 214 pages of tender documents read, 68 binding clauses pulled out, 9,485 building parts loaded, 14 places found where a duct and a beam sit in the same cubic metre, deepest one 38mm, all of them fixed, 3,318 lines of quantities priced and the package encrypted and uploaded before the 17:00 cutoff this is Graph Engineering: the job gets cut into small nodes, one narrow task each, wired so that one node's output is the next node's input, and any node is allowed to stop the whole run. it turns a model that answers you into a machine that finishes the job: - give every node one job and one output. a node doing two things fails at both and you cannot tell which one broke - put the cheapest rejection first. his qualification node reads clause 7.4, foreign-owned firms barred, and ends the run four seconds in, before anything expensive touches the model - what moves between nodes is a file. the model travels as a model, the quantities as a table, the price as a number - build exactly one loop: the checker finds 14 collisions, the fixer drops the duct 550mm, the checker runs again, and nothing moves on until the count is zero - cap that loop, or a graph will grind on three impossible clashes until the deadline passes - keep one node whose only job is to say no, and give it authority over everything above it - log each node's output on its own, because when the price comes out wrong you need to know which node believed the wrong thing - run the expensive nodes last, always the catch is that a graph is an extremely confident machine: point it at an outdated rate book and it prices an entire hospital off it without a single node noticing, because no node is asked to doubt the input, only to process it so the nodes that earn their keep are the ones that reject, and almost nobody builds those first bookmark this, the full build with all nine nodes and what each one hands to the next is written out in the article ↓

Argona

38,189 görüntüleme • 1 ay önce

Most people think their AI is not smart enough. It is smart enough. It just knows nothing about you. Every chat starts from zero. You paste the doc. You explain the project. You explain the same project again tomorrow. I finally fixed that part: 👇 I have been running Littlebird for a while now. The idea is simple. It is a desktop app for Mac and Windows that reads the text on your active window and sits in on your calls. So it already knows what you have been working on before you ask it anything. You do not brief it. You just ask. What that looks like in a normal day: 1/ Chat that already has the context No pasting. I ask what changed in a brief last week and it answers from the actual document that was open on my screen. 2/ Meeting Notes that write themselves It transcribes the call, then hands me the decisions and the action items. I get to stay in the conversation instead of typing through it. 3/ Routines that run on a schedule A morning briefing. A weekly summary of what I actually shipped. It shows up on its own. 4/ Hummingbird for the small stuff It appears right where you are working, so a quick question does not cost you a window switch and ten minutes of drift. The point is not the notes. Plenty of apps take notes. The point is that I stopped re-explaining my own work to a machine fifteen times a day. Free plan if you want to test the idea before deciding. Link is in the first comment, along with a discount for new users.

Most people think their AI is not smart enough. It is smart enough. It just knows nothing about you. Every chat starts from zero. You paste the doc. You explain the project. You explain the same project again tomorrow. I finally fixed that part: 👇 I have been running Littlebird for a while now. The idea is simple. It is a desktop app for Mac and Windows that reads the text on your active window and sits in on your calls. So it already knows what you have been working on before you ask it anything. You do not brief it. You just ask. What that looks like in a normal day: 1/ Chat that already has the context No pasting. I ask what changed in a brief last week and it answers from the actual document that was open on my screen. 2/ Meeting Notes that write themselves It transcribes the call, then hands me the decisions and the action items. I get to stay in the conversation instead of typing through it. 3/ Routines that run on a schedule A morning briefing. A weekly summary of what I actually shipped. It shows up on its own. 4/ Hummingbird for the small stuff It appears right where you are working, so a quick question does not cost you a window switch and ten minutes of drift. The point is not the notes. Plenty of apps take notes. The point is that I stopped re-explaining my own work to a machine fifteen times a day. Free plan if you want to test the idea before deciding. Link is in the first comment, along with a discount for new users.

Mushfiq Sajib

66,683 görüntüleme • 8 gün önce

i gave Ox Alpha the main menu from ECHO and asked it to recreate it only ONE (1) prompt was enough to build this, using the reference this is one of the most beautiful output i've ever had from any model the fact that it's from a stealth model is mindblowing

i gave Ox Alpha the main menu from ECHO and asked it to recreate it only ONE (1) prompt was enough to build this, using the reference this is one of the most beautiful output i've ever had from any model the fact that it's from a stealth model is mindblowing

filipe

31,171 görüntüleme • 22 gün önce

We took a 30B model and split it in two to write tokens in parallel instead of one at a time. Introducing Nemotron-Labs-TwoTower: a diffusion language model from NVIDIA Research adapted from Nemotron-3-Nano-30B-A3B. Here’s how it works: one half holds the context, the other writes the tokens, with both reusing the pretrained model instead of training a new one from scratch. We found it kept 98.7% of the original model’s quality at 2.42× faster generation.

We took a 30B model and split it in two to write tokens in parallel instead of one at a time. Introducing Nemotron-Labs-TwoTower: a diffusion language model from NVIDIA Research adapted from Nemotron-3-Nano-30B-A3B. Here’s how it works: one half holds the context, the other writes the tokens, with both reusing the pretrained model instead of training a new one from scratch. We found it kept 98.7% of the original model’s quality at 2.42× faster generation.

NVIDIA AI

762,265 görüntüleme • 2 ay önce

the thing you rent for $200 a month just became something you can own for $1,700 once but the money is not even the real story for the first time a 200 billion parameter model is not in a datacenter, it is sitting on a desk the cloud spent years convincing you a model this size needed their servers, their meter, their monthly bill people are stacking four subscriptions into a $440 a month bill to rent what one box this size now owns outright it needed a box the size of a book the moment the model moves from their datacenter to your desk, the whole game changes it stops being about who has the best AI it becomes about who ships it on every desk the cloud told you this needed a datacenter it needed a desk i did the full math on what this kills in the article below

the thing you rent for $200 a month just became something you can own for $1,700 once but the money is not even the real story for the first time a 200 billion parameter model is not in a datacenter, it is sitting on a desk the cloud spent years convincing you a model this size needed their servers, their meter, their monthly bill people are stacking four subscriptions into a $440 a month bill to rent what one box this size now owns outright it needed a box the size of a book the moment the model moves from their datacenter to your desk, the whole game changes it stops being about who has the best AI it becomes about who ships it on every desk the cloud told you this needed a datacenter it needed a desk i did the full math on what this kills in the article below

John Doe

26,096 görüntüleme • 3 ay önce

Big win for open-source LLMs! DeepSeek V4 Pro holds the top open-weights score on SWE-bench Verified, in the GPT-5.5 range. GLM 5.2 leads the open-weight intelligence index and sits near the closed frontier on long-horizon coding. But this leaderboard number is a weak proxy for real performance. It comes from one task set, run through one harness, served at one precision. The same weights can even score differently across providers, since many hosts quantize activations to fp8 and drift the model off its reference weights. Real performance is determined based on whether a model can read a repo, make coordinated edits across files, run the tests, and recover when one breaks. By that measure, the top open models hold up, but only inside the right harness. The teams that actually put DeepSeek V4 into production pipelines as a frontier substitute got there through the harness they built around the model, not by picking a stronger model. If you want to see this in practice, Cline (64k+ stars) has actually built that harness around open models, tuned so they run at production quality. And it's tuned so that these LLMs can run at production quality, with plan and act modes, checkpoints, and terminal feedback. ClinePass is the new access layer on top of it. It runs a curated set of those models inside Cline, narrowed to the ones tested for coding-agent use, with 2 to 5x the standard rate limits and no separate provider accounts, keys, or billing to track. The video below shows the setup, and I worked with the team to put this together. It runs alongside custom keys and local models as well, not in place of them.

Big win for open-source LLMs! DeepSeek V4 Pro holds the top open-weights score on SWE-bench Verified, in the GPT-5.5 range. GLM 5.2 leads the open-weight intelligence index and sits near the closed frontier on long-horizon coding. But this leaderboard number is a weak proxy for real performance. It comes from one task set, run through one harness, served at one precision. The same weights can even score differently across providers, since many hosts quantize activations to fp8 and drift the model off its reference weights. Real performance is determined based on whether a model can read a repo, make coordinated edits across files, run the tests, and recover when one breaks. By that measure, the top open models hold up, but only inside the right harness. The teams that actually put DeepSeek V4 into production pipelines as a frontier substitute got there through the harness they built around the model, not by picking a stronger model. If you want to see this in practice, Cline (64k+ stars) has actually built that harness around open models, tuned so they run at production quality. And it's tuned so that these LLMs can run at production quality, with plan and act modes, checkpoints, and terminal feedback. ClinePass is the new access layer on top of it. It runs a curated set of those models inside Cline, narrowed to the ones tested for coding-agent use, with 2 to 5x the standard rate limits and no separate provider accounts, keys, or billing to track. The video below shows the setup, and I worked with the team to put this together. It runs alongside custom keys and local models as well, not in place of them.

Avi Chawla

44,124 görüntüleme • 2 ay önce

watch anon. a 27b model thinking out loud on gtx 1660 super, i asked bonsai what model it is and it reasoned through the whole answer before it spoke. 20 tokens a second, gpu pinned at 125 watts, 4.25 of 6 gigs used. bonsai is a 1bit quant of qwen 3.6 27b, the king of the 3090 crushed to 3.5gb. turns out the hardware was in your drawer the whole time. you just needed the quant to catch up.

watch anon. a 27b model thinking out loud on gtx 1660 super, i asked bonsai what model it is and it reasoned through the whole answer before it spoke. 20 tokens a second, gpu pinned at 125 watts, 4.25 of 6 gigs used. bonsai is a 1bit quant of qwen 3.6 27b, the king of the 3090 crushed to 3.5gb. turns out the hardware was in your drawer the whole time. you just needed the quant to catch up.

Sudo su

26,133 görüntüleme • 1 ay önce

MiniMax M3 just dropped — their first natively multimodal model. So I ran it through my form-filling test. (The model has to place each element at the right pixel position on a blank form image, not type into a field.) Verdict: it got everything on the paper. > Name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code, all there. > Best character spacing I've seen yet: it actually calculates the gap between each character, clean across the DOB and number boxes > A few fields slightly misaligned, but every piece of data made it onto the form The reasoning chain is the interesting part: it does the easy fields first, then works into the tight one-char-per-box fields, reasoning through y-coordinates, baselines, and label clearance in obsessive detail. The cost: 40:33 and 126.7k output tokens. That's a long think — but it's MiniMax's first multimodal model, and it nailed the content.

MiniMax M3 just dropped — their first natively multimodal model. So I ran it through my form-filling test. (The model has to place each element at the right pixel position on a blank form image, not type into a field.) Verdict: it got everything on the paper. > Name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code, all there. > Best character spacing I've seen yet: it actually calculates the gap between each character, clean across the DOB and number boxes > A few fields slightly misaligned, but every piece of data made it onto the form The reasoning chain is the interesting part: it does the easy fields first, then works into the tight one-char-per-box fields, reasoning through y-coordinates, baselines, and label clearance in obsessive detail. The cost: 40:33 and 126.7k output tokens. That's a long think — but it's MiniMax's first multimodal model, and it nailed the content.

stevibe

27,383 görüntüleme • 3 ay önce

There is a room in Málaga that was built to be the closest thing on earth to standing inside heaven. It is called the camarín of the Virgin of Victory, and it is hidden at the top of a tower inside the Santuario de la Victoria. To reach it, you climb and the ascent is the entire point... The building you are climbing through was completed in 1700, and it was designed as a single argument made in stone. At the bottom lies a crypt: a black chamber crowded with white plaster skeletons, a meditation on death and the brevity of life. From there a staircase rises, and as you climb it the light grows stronger and the imagery changes from bones to saints. The architects of the time understood this ascent as the soul's own journey, the dark crypt as the stage of penitence, the staircase as the stage of spiritual progress, and the room at the very top as the final stage: the union of the soul with the divine. That room at the top is the camarín, and its dome is one of the most extraordinary interiors in Spain... Every surface is covered in white and gold plasterwork. There is no empty space anywhere. The Baroque called this horror vacui, the horror of the void: the conviction that a space meant to represent heaven should not contain a single bare patch of stone. Out of that plasterwork emerge angels, flowers, birds, and mirrors. The mirrors are not decoration alone. They catch the light pouring in through the windows of the drum and throw it around the chamber, so that the gold seems to move and the whole room appears to shimmer and breathe. This wonder was built by people who believed that if you wanted to show a human being what heaven might feel like, you did not describe it to them. You built a room, and you let them climb into it... -- -- -- If you enjoyed this, I write a weekly newsletter read by over 50,000 people who love rediscovering the beauty of the past. You can join us here: If you'd like to support my work, a paid subscription is what makes it possible.

There is a room in Málaga that was built to be the closest thing on earth to standing inside heaven. It is called the camarín of the Virgin of Victory, and it is hidden at the top of a tower inside the Santuario de la Victoria. To reach it, you climb and the ascent is the entire point... The building you are climbing through was completed in 1700, and it was designed as a single argument made in stone. At the bottom lies a crypt: a black chamber crowded with white plaster skeletons, a meditation on death and the brevity of life. From there a staircase rises, and as you climb it the light grows stronger and the imagery changes from bones to saints. The architects of the time understood this ascent as the soul's own journey, the dark crypt as the stage of penitence, the staircase as the stage of spiritual progress, and the room at the very top as the final stage: the union of the soul with the divine. That room at the top is the camarín, and its dome is one of the most extraordinary interiors in Spain... Every surface is covered in white and gold plasterwork. There is no empty space anywhere. The Baroque called this horror vacui, the horror of the void: the conviction that a space meant to represent heaven should not contain a single bare patch of stone. Out of that plasterwork emerge angels, flowers, birds, and mirrors. The mirrors are not decoration alone. They catch the light pouring in through the windows of the drum and throw it around the chamber, so that the gold seems to move and the whole room appears to shimmer and breathe. This wonder was built by people who believed that if you wanted to show a human being what heaven might feel like, you did not describe it to them. You built a room, and you let them climb into it... -- -- -- If you enjoyed this, I write a weekly newsletter read by over 50,000 people who love rediscovering the beauty of the past. You can join us here: If you'd like to support my work, a paid subscription is what makes it possible.

James Lucas

69,389 görüntüleme • 3 ay önce

encrypt every chat with this open sourced repo 18 year old just killed the encryption problem it's called conversation-steganography you type a secret and a local AI buries it inside a dead boring chat about coffee, etc. then you send that through any normal messenger your friend runs the same tool and the small talk unfolds back into the real words > the secret is locked with AES-SIV before it ever becomes text > the cover message is written by a model on your own machine so nothing touches the cloud > it rides through whatsapp telegram imessage even plain email > every message is cryptographically chained so tampering shows instantly whatsapp sees you planning dinner your isp logs small talk the real sentence was never in the data at all quietly one of the most dangerous repos i have seen all month repo in reply

encrypt every chat with this open sourced repo 18 year old just killed the encryption problem it's called conversation-steganography you type a secret and a local AI buries it inside a dead boring chat about coffee, etc. then you send that through any normal messenger your friend runs the same tool and the small talk unfolds back into the real words > the secret is locked with AES-SIV before it ever becomes text > the cover message is written by a model on your own machine so nothing touches the cloud > it rides through whatsapp telegram imessage even plain email > every message is cryptographically chained so tampering shows instantly whatsapp sees you planning dinner your isp logs small talk the real sentence was never in the data at all quietly one of the most dangerous repos i have seen all month repo in reply

savip.

83,651 görüntüleme • 1 ay önce

SOMEONE JUST VIBECODED A PROGRAM THAT DOWNLOADS ANY WEBSITE ON THE INTERNET every $35,000 agency site is now yours to rip clean for $0 fully open source, save it before it's gone you paste one url and it does the rest: → crawls the entire site → grabs every file: html, css, js, images, even the fonts → rewrites every link to run fully offline → packs it all into a downloadable zip runs entirely in the browser, nothing to install or sign up for point it at any site you admire and the whole front-end is yours: → clone it → archive it before it dies → crack open exactly how it was built the thing agencies gatekeep for five figures is now a 10-second download bookmark this one

SOMEONE JUST VIBECODED A PROGRAM THAT DOWNLOADS ANY WEBSITE ON THE INTERNET every $35,000 agency site is now yours to rip clean for $0 fully open source, save it before it's gone you paste one url and it does the rest: → crawls the entire site → grabs every file: html, css, js, images, even the fonts → rewrites every link to run fully offline → packs it all into a downloadable zip runs entirely in the browser, nothing to install or sign up for point it at any site you admire and the whole front-end is yours: → clone it → archive it before it dies → crack open exactly how it was built the thing agencies gatekeep for five figures is now a 10-second download bookmark this one

Argona

126,379 görüntüleme • 2 ay önce

THIS IS F**KING DANGEROUS THREE OPEN SOURCE AI CYBERSECURITY STACK MOST PEOPLE DONT KNOW EXISTS METATRON: point it at a target, it recon's, analyzes, and reports on its own —> fully OFFLINE , zero API keys. NUCLEI : describe a vulnerability in plain English, it writes and runs the actual detection template for you. GARAK BY (NVIDIA): throws thousands of known prompt injection and jailbreak attempts at an LLM until one breaks through. Recon —> detection —> and AI red teaming three layers of the same problem, all open source. Link to access the REPO’S below

THIS IS F**KING DANGEROUS THREE OPEN SOURCE AI CYBERSECURITY STACK MOST PEOPLE DONT KNOW EXISTS METATRON: point it at a target, it recon's, analyzes, and reports on its own —> fully OFFLINE , zero API keys. NUCLEI : describe a vulnerability in plain English, it writes and runs the actual detection template for you. GARAK BY (NVIDIA): throws thousands of known prompt injection and jailbreak attempts at an LLM until one breaks through. Recon —> detection —> and AI red teaming three layers of the same problem, all open source. Link to access the REPO’S below

MAX

104,453 görüntüleme • 25 gün önce

If you have an RTX 3090 or 4090, Mia just shipped you a free massive upgrade in both speed and intelligence. I will explain to you why this will make your Qwen 3.8 27B on your card, even better, and my flags for running it. Qwen3.8-27B, EXL3 3.5bpw, DFlash2 speculative decode, RTX 4090. Single stream. The kit is from MiaAI-Lab, EXL3 is turboderp's format. I re-measured everything on my own card because the my first benchmarks seemed off. It turns out it really does run much faster. WHY EXL3 IS A DIFFERENT ANIMAL The old way (Q4_K_M) rounds each weight to the nearest 4-bit value independently. Every weight introduces its own rounding error. Those errors accumulate across millions of weights and causes drift (Slightly dumber). EXL3 is a fundamentally different compression algorithm. Instead of rounding each weight on its own, it encodes the entire weight vector as a path through a constrained codebook and spreads the rounding error across dimensions using a Hadamard transform. The result is that at the same bits per weight, more of the original model's intelligence is preserved. The important part is this CAN ACTUALLY BE MEASURED. The cleanest way to see that is KL divergence against a high-precision teacher. Lower means the quantized model thinks more like the original. On the malaiwah independent teacher-logit panel for GLM-5.3-Flash: EXL3 4bpw: 0.0246 nats Official FP8: 0.0206 nats NVFP4: 0.0605 nats EXL3 sits 0.004 nats behind native FP8 at half the size. NVFP4 at higher bit width is 2.5x further from the teacher. That panel is GLM-5.3-Flash, not Qwen 3.8. Cited as the mechanism, not as this run's data. But the point stands: EXL3 is not just smaller, it is smarter per bit than the formats most people are running. WHAT I MEASURED I first measure 108 tok/s from a single run. After that number looked too good to be true. I reran it. It looks like after a warm up, the numbers are even better. Basically, like people long thought, the RTX 3090 and RTX 4090 are actually superb AI computer cards. Hence why NVIDIA stopped shipping them with NVLINK since the 4090. Short context ceiling (~2k in, 1016-token output, TTFT-separated): 135, 138, 153, 174, 133, 129 tok/s across 6 runs. Sustained longform (2040-token essay): 105.4, 94.5, 98.4 tok/s Short answer (504 tokens): 93.2 tok/s The honest shape: ~130-150 tok/s at short context is the ceiling, ~94-105 sustained on longform. The ceiling matters because that is what people feel in chat. The old dense Q4_K_M on llama.cpp ran ~37 tok/s on this same card. (No MTP), with MTP about 60 tok/s Sustained is roughly 2.5-3x. Ceiling is closer to 4x. Same model, different quantization and engine. The multiplier comes from EXL3, the ExLlamaV2 engine, and DFlash2 together. CONCURRENCY IS A RTX 4090 LANE. Just like the old config on the 4090, the 24gb vram, means a long context can only hold one stream, and running concurrency requires to lower context length, because it runs fast it sort of makes up for it by being faster than slower GPU chips. CONTEXT LADDER The recipe doc measured prefill. I re-ran it with TTFT separated from decode, because decode is what you actually feel after the first token. ~5k in: decode 140 tok/s (TTFT 0.5s), needle HIT ~18k in: decode 85 tok/s (TTFT 0.3s*), needle HIT ~73k in: decode 28 tok/s (TTFT 1.8s), needle HIT ~146k in: decode 16 tok/s (TTFT 2.3s), needle HIT (*0.3s at 18k is a prefix-cache hit from the paired pass. Cold prefill for reference: ~2,020 tok/s at 17k falling to ~508 at 153k.) Needle hit at every depth, mine and the original 7/7. Retrieval is intact at max context. Speed is not: decode falls ~9x from short to max. Past ~50k tokens this stops being a chat tool and becomes a batch tool. At 146k it works, but nobody is typing interactively against 16 tok/s. WHERE IT BROKE The model's native context is 262k. The README says DFlash2 fits ~220k on a 24GB card. My 200,704-token attempt failed with insufficient VRAM. Dropped to 168,960 and it booted. The real ceiling is somewhere between 168,960 and 200,704 and I never tested that gap. I jumped to a value that worked and called it done. That is ~32k tokens of context I left on the table. One thing the numbers taught me: JSON tokenizes at ~1.5 chars/token, prose at ~3.9. "150k tokens of JSON" needs ~2.7x more filler than the same estimate in prose. Size by real tokens, not estimates. THE UPGRADE If you own a 4090 and you are running Q4_K_M on llama.cpp, you are leaving a good bit of speed and measurable intelligence on the table. The same model, on the same card, with a better quantization and engine, goes from 60 tok/s to 130-150 at short context and 94-105 sustained. The model also thinks closer to the original because EXL3 preserves more of the output distribution per bit than the old rounding method. The recipe is in the first reply. Everything above came from one 4090 and one afternoon of re-measuring. The decode ladder especially needs independent numbers. If your card gets different falloff, that is worth knowing. Recipe and flags/ findings in reply 👇

If you have an RTX 3090 or 4090, Mia just shipped you a free massive upgrade in both speed and intelligence. I will explain to you why this will make your Qwen 3.8 27B on your card, even better, and my flags for running it. Qwen3.8-27B, EXL3 3.5bpw, DFlash2 speculative decode, RTX 4090. Single stream. The kit is from MiaAI-Lab, EXL3 is turboderp's format. I re-measured everything on my own card because the my first benchmarks seemed off. It turns out it really does run much faster. WHY EXL3 IS A DIFFERENT ANIMAL The old way (Q4_K_M) rounds each weight to the nearest 4-bit value independently. Every weight introduces its own rounding error. Those errors accumulate across millions of weights and causes drift (Slightly dumber). EXL3 is a fundamentally different compression algorithm. Instead of rounding each weight on its own, it encodes the entire weight vector as a path through a constrained codebook and spreads the rounding error across dimensions using a Hadamard transform. The result is that at the same bits per weight, more of the original model's intelligence is preserved. The important part is this CAN ACTUALLY BE MEASURED. The cleanest way to see that is KL divergence against a high-precision teacher. Lower means the quantized model thinks more like the original. On the malaiwah independent teacher-logit panel for GLM-5.3-Flash: EXL3 4bpw: 0.0246 nats Official FP8: 0.0206 nats NVFP4: 0.0605 nats EXL3 sits 0.004 nats behind native FP8 at half the size. NVFP4 at higher bit width is 2.5x further from the teacher. That panel is GLM-5.3-Flash, not Qwen 3.8. Cited as the mechanism, not as this run's data. But the point stands: EXL3 is not just smaller, it is smarter per bit than the formats most people are running. WHAT I MEASURED I first measure 108 tok/s from a single run. After that number looked too good to be true. I reran it. It looks like after a warm up, the numbers are even better. Basically, like people long thought, the RTX 3090 and RTX 4090 are actually superb AI computer cards. Hence why NVIDIA stopped shipping them with NVLINK since the 4090. Short context ceiling (~2k in, 1016-token output, TTFT-separated): 135, 138, 153, 174, 133, 129 tok/s across 6 runs. Sustained longform (2040-token essay): 105.4, 94.5, 98.4 tok/s Short answer (504 tokens): 93.2 tok/s The honest shape: ~130-150 tok/s at short context is the ceiling, ~94-105 sustained on longform. The ceiling matters because that is what people feel in chat. The old dense Q4_K_M on llama.cpp ran ~37 tok/s on this same card. (No MTP), with MTP about 60 tok/s Sustained is roughly 2.5-3x. Ceiling is closer to 4x. Same model, different quantization and engine. The multiplier comes from EXL3, the ExLlamaV2 engine, and DFlash2 together. CONCURRENCY IS A RTX 4090 LANE. Just like the old config on the 4090, the 24gb vram, means a long context can only hold one stream, and running concurrency requires to lower context length, because it runs fast it sort of makes up for it by being faster than slower GPU chips. CONTEXT LADDER The recipe doc measured prefill. I re-ran it with TTFT separated from decode, because decode is what you actually feel after the first token. ~5k in: decode 140 tok/s (TTFT 0.5s), needle HIT ~18k in: decode 85 tok/s (TTFT 0.3s), needle HIT ~73k in: decode 28 tok/s (TTFT 1.8s), needle HIT ~146k in: decode 16 tok/s (TTFT 2.3s), needle HIT (0.3s at 18k is a prefix-cache hit from the paired pass. Cold prefill for reference: ~2,020 tok/s at 17k falling to ~508 at 153k.) Needle hit at every depth, mine and the original 7/7. Retrieval is intact at max context. Speed is not: decode falls ~9x from short to max. Past ~50k tokens this stops being a chat tool and becomes a batch tool. At 146k it works, but nobody is typing interactively against 16 tok/s. WHERE IT BROKE The model's native context is 262k. The README says DFlash2 fits ~220k on a 24GB card. My 200,704-token attempt failed with insufficient VRAM. Dropped to 168,960 and it booted. The real ceiling is somewhere between 168,960 and 200,704 and I never tested that gap. I jumped to a value that worked and called it done. That is ~32k tokens of context I left on the table. One thing the numbers taught me: JSON tokenizes at ~1.5 chars/token, prose at ~3.9. "150k tokens of JSON" needs ~2.7x more filler than the same estimate in prose. Size by real tokens, not estimates. THE UPGRADE If you own a 4090 and you are running Q4_K_M on llama.cpp, you are leaving a good bit of speed and measurable intelligence on the table. The same model, on the same card, with a better quantization and engine, goes from 60 tok/s to 130-150 at short context and 94-105 sustained. The model also thinks closer to the original because EXL3 preserves more of the output distribution per bit than the old rounding method. The recipe is in the first reply. Everything above came from one 4090 and one afternoon of re-measuring. The decode ladder especially needs independent numbers. If your card gets different falloff, that is worth knowing. Recipe and flags/ findings in reply 👇

Yume_X

38,685 görüntüleme • 14 gün önce

Okay, I am fairly confident in my hypothesis now. The secret sauce behind Qwen 3.8 27B becomes almost immediately evident during testing. It is not the training data. In fact, I doubt any SFT was involved at all. The model was simply allowed GRPO with a more liberal reasoning context length. As I have explained to you before, distillation through GRPO is powerful, and in certain contexts vastly superior to K/L divergence distillation. Qwen used the outputs of Qwen 3.8 Max as the reinforcement learning objective for Qwen 3.8 27B and told the model reason as much as it has to in order to match the output. Remember, Intelligence is solution, not compression. This is like nesting the output of a function as an additional input to itself, and it is absolutely brilliant.

Okay, I am fairly confident in my hypothesis now. The secret sauce behind Qwen 3.8 27B becomes almost immediately evident during testing. It is not the training data. In fact, I doubt any SFT was involved at all. The model was simply allowed GRPO with a more liberal reasoning context length. As I have explained to you before, distillation through GRPO is powerful, and in certain contexts vastly superior to K/L divergence distillation. Qwen used the outputs of Qwen 3.8 Max as the reinforcement learning objective for Qwen 3.8 27B and told the model reason as much as it has to in order to match the output. Remember, Intelligence is solution, not compression. This is like nesting the output of a function as an additional input to itself, and it is absolutely brilliant.

Aesther

122,746 görüntüleme • 1 ay önce