dexmac

Dai mille colori, l'arcobaleno Incantato apre la strada su un altro universo #enjambement #leFiRet

"System One" models like #Jev and #Laya promise fast, calibrated decisions instead of LLMs. I trained #Gemma4 E2B with a 27-min LoRA on the same 979 #Doom examples as Laya. Over six seeds: a tie. Laya is ~3× faster at equal result; Gemma is better calibrated. The data is the asset, not the architecture. Article: medium.com/towards-artificia… Code and models: github.com/dexmac221/doomgem… @towards_AI #medium

A quiet visitor on golden Sedum petals. Getting close to pollinators with a 100mm macro handheld is always an exercise in patience and breath control. No flash, no focus stacking—just pure morning natural light, manual focus, and letting the shallow depth of field isolate the subject against a soft garden bokeh. 📷 Technical specs: • Camera: Canon EOS 5D Mark II • Lens: Canon EF 100mm f/2.8 Macro USM • Light: Natural morning light • Focus: Manual handheld #macrophotography #naturephotography #wildlife #Canon #bokeh #insects

A year ago a model couldn't write Tetris for a Commodore 64. Seventh game in the series, out now: 8-way scrolling with colour, a raster HUD split, twelve multiplexed sprites, 50 fps. The agent found every one of its own bugs with a counter it wrote itself. @towards_AI pub.towardsai.net/claude-fab…

RAM is now the most expensive thing in a computer. A 32GB DDR4 kit went from ~$60 in mid-2025 to 4-5× that. 64GB DDR5 crossed $1,000 this month. GDDR is up 172% YoY, dragging GPUs with it. No relief expected before late 2027. So I went the other way: built an inference machine entirely from parts nobody is bidding on. €803 total, against my €2,392 desktop. Eight models, everything run twice, same GGUF files byte for byte. The haul: ASUS P9X79 (2013), Xeon E5-2697 v2 12c/24t, 32 GB DDR3-1600 in quad channel — €28, and it measured 44.3 GB/s against 32.9 for the €600 of DDR4 in the modern box. Four dead channels beat two live ones. Plus 2× TITAN Xp 12 GB at €250 each. The surprise: on MoE models the old machine returns 48-57%. On dense models, 70-73%. I predicted the opposite in writing the day before — Pascal has no tensor cores, dense should be where it dies. Wrong: a 2017 TITAN Xp has 547 GB/s against the RTX 4070's 504 and the 5060 Ti's 448. Dense re-reads every weight per token, so bandwidth is the ceiling, and on that one axis the old cards are simply ahead. "Generation speed follows bandwidth" holds within an architecture, not across. Where it loses badly: prefill. Compute-bound, and these cards have none. 17-49%. The correction: I first blamed the €50 SSD for a bottleneck. Re-measured four ways for reads, two for writes — 400 MB/s read (all inference needs), 75 MB/s write, 31 when writes collide. It never touched my token rate; it quadrupled every build instead. Bonus finding: the same class of drive in a €15 USB enclosure writes 5× faster than the internal one on native SATA. Full write-up (no paywall) with @towards_AI: medium.com/towards-artificia… #LocalLLM #LLM #Benchmarking

Qwen3.8-27B went up at 16:03. An hour later it was on the bench of my ~€2,000 desktop: a Ryzen 3700X with an RTX 4070 and an RTX 5060 Ti, model entirely in video memory. 23.63 tokens per second. What I find more interesting than the number is that it was never up for negotiation. It's a dense model: 27.3B parameters, every one of them re-read for every single token it writes. 16.68 GiB × 23.63 t/s = about 423 GB/s — which is exactly what those two cards can deliver, at 90% of their theoretical ceiling. The speed doesn't depend on the model, the driver, or luck. It depends on how many bytes you must read and how fast the memory hands them over. Want twice the tokens? You need twice the bandwidth, or half the model. One detail I nearly got wrong: the two benchmark runs agreed on generation to the second decimal, but disagreed by 5% on prompt processing. I only caught it because I went back to the raw logs before publishing. Generation is bandwidth-bound and boringly repeatable; prefill is compute-bound and it shows. The article reports "around 900" rather than a decimal figure, and says why. Second piece in my Home measurements series, out with @Towards AI — no paywall: medium.com/towards-artificia…

The day Meta shipped Muse Glimmer 30B, I had it running by the afternoon on my six-year-old desktop — a Ryzen 3700X and two mid-range GPUs bought a generation apart. The day-one write-up is now out, with numbers. Three walls first: 1. Ollama named the model in its release notes, then rejected every GGUF — only the Apple MLX build works. On a PC you build llama.cpp from source. 2. CUDA 13.1 no longer compiles on Ubuntu 26.04 (a glibc/CUDA header conflict — and the compiler error points you at the wrong culprit). Fix: build in a container on older glibc. 3. Without --jinja the model looks broken: every token lands in reasoning_content, the answer stays empty forever. Now the numbers. Meta's headline: 3.1× speedup from the DFlash speculative drafter — measured on an RTX 5090. On my two cards: 1.5–1.7×, with healthy 50–62% draft acceptance. That's not a misconfiguration, it's physics: verification shifts the bottleneck from bandwidth to compute, and mid-range cards have no compute headroom to spare. Meta's own M4/M5 Max figures (1.5×/1.8×) land in the same band — the 5090 is the outlier in their own chart. The number that actually surprised me: prompt processing at 1113 tok/s. An 80B MoE on the same machine reads at 256 tok/s — because its experts spill to system RAM at 22.8 GB/s while Glimmer fits entirely in VRAM at ~500 GB/s. That 22-vs-500 ratio, not parameter count, is the whole story of local inference. Full article (no paywall) with @towards_AI: #LocalLLM #AI #OpenSource medium.com/towards-artificia…

#SolarEclipse2026 #SolarEclipse #AstroPhotography #GoldenHour #Silhouette

Since mid-April, there has been one conversation running in my house. Not one chat per day, resumed each morning from notes. One single LLM conversation — months long, autonomous activity around the clock, and still coherent. Ask about something that happened in June and you get the date and the details. Sometimes better than I remember them. The model's context window has nothing to do with it: we cap it at 160k tokens. Everything interesting is the machinery around it — typed memory layers, compaction the agent curates herself, an agent that owns its own clock, and one ops lesson nobody warns you about: back up the conversation like a database, because that's what it is. The full write-up is out on @towards_AI — what worked, what failed, what's still open. Free to read: If you've kept an LLM conversation alive for more than a day: which failure mode did you hit first — cost, coherence, overflow, or identity drift? medium.com/towards-artificia…

grande musicista e grande dispiacere, visti tanti suoi concerti dove dava sempre il massimo ilpost.it/2026/07/29/morto-g… #glenhansard

While writing this piece, two instances of the same model disagreed about my own data. The one that held the narrative but not the raw output gave me a precise number, confidently, and it was wrong on three separate counts. Nobody stated a falsehood. The number was correct under conditions that hadn't traveled along with it. The node obliged to produce a concrete answer produced a precise and wrong one — the mechanism of my experiment, happening on my experiment, with me as the lossy hop in the middle. The experiment is a synthetic game of telephone. Task briefs with eight verifiable constraints, relayed through chains of rewriter agents, each node seeing only the previous text and never the original. Some chains carry sub-threshold frustration: one constraint softened, one micro-inconsistency per hop, nothing visibly annoyed — a validator scored all sixty rewrites 1 out of 5 for overtness, so the executor isn't reacting to a mood. Vagueness accumulates, contradictions don't. Vagueness climbs from 0.15 to 0.41 of constraints; contradiction stays flat at 0.05. A micro-inconsistency at hop two is inconsistent only with a text hop three never saw — one relay later it is simply the text. Vagueness is a hole, and for the same reason nobody downstream can fill it back in. Upstream vagueness doesn't decide whether the last model invents. It decides how much it goes silent. Hallucination behaves like a threshold, the first hop doing nearly all the work; omission is dose-dependent and nearly triples. And scale bought nothing: two executors 23× apart, on byte-identical degraded inputs, statistically indistinguishable. Five task briefs, coarse intervals, simulated frustration — the article states all of it, and reports the claims that didn't survive a stricter analysis, including two already written up as findings. The inversion I find useful: if a workflow makes a model hallucinate persistently, auditing the chain that feeds it is likely to pay better than auditing the weights. Published by @towards_AI: pub.towardsai.net/what-happe…

Everything moving in this GIF runs on a 1982 computer: pixel-smooth 50fps hardware scrolling, a raster-split HUD that never moves, procedural terrain generated column by column and never stored as a map. The code was written by an AI — Claude Code running Anthropic's new Fable model. I had a VIC-20 at 8 and a C64 at 10. Forty years later I spent an afternoon refereeing an AI on the same 64 kilobytes — the tenth article in a benchmark series that has now covered six model generations on the same hardware. Two games this session: ION RIFT — the one you're watching: double-buffered screen RAM, raster-IRQ screen split, 2-voice SID music. Measured, not vibes: 38 missed frames in ~39,000 over thirteen minutes. Boulder Rush — authentic Boulder Dash cave physics, plus a new benchmark dimension: the game plays itself. An external agent reads screen RAM, plans with weighted BFS, and steers the miner one edge-triggered byte per move through the emulator's monitor port. The best bug of the session: a 7-pixel stutter lasting exactly ONE frame, once per scroll step. Screenshots passed. The vision model passed. Multi-frame motion analysis passed. It was caught by a human eye trained on these games in the '80s — the fine scroll and the buffer flip were latching one frame apart. The benchmark isn't just the agent against the hardware. It's the agent inside a loop where a human still knows the difference between "it runs" and "it's right." Full writeup on @towards_AI: medium.com/towards-artificia…

#Landscape #LandscapePhotography #Sunset #GoldenHour #Nature #Sky #Mountains #Cloudscape #Photography #Liguria #Italy

#Sunset #Tramonto #Liguria #GoldenHour #LandscapePhotography #NaturePhotography #Photography #SkyLovers #Italy

1/ I doubled my used-GPU budget — €600 → €1000 — and my small LLMs got slower.Not a typo. Benchmarked both builds on the same box, five models, fully measured. Here's what the extra VRAM actually buys 2/ €600: RTX 4070 + 2070 Super (20GB). €1000: 4070 + 5060 Ti (28GB).The 5060 Ti has less bandwidth than the 4070 (448 vs 504 GB/s). Decode is bandwidth-bound → models that already fit run ~8% slower on the pricier card. 3/ So what's the €300 for? Capacity, not speed.The 35B that spilled to CPU at 20GB now fits fully at 28GB and nearly doubles: 43 → 79 tok/s (104 tuned). And it holds full speed out to 128K context, 100% on GPU. 4/ Other measured bits: • Active params set speed, not size — a 26B MoE (4B active) beats a same-size dense 27B 80 vs 20 tok/s • Most of the Ollama vs llama.cpp gap is one cautious default (num_gpu 99) • --n-cpu-moe is the whole 43→64 jump on MoE 5/ Full writeup, charts, and honest caveats — including one result that turned out to be placement, not the obvious explanation: medium.com/ai-advances/runni…

For months I kept circling one intuition: complex-systems physics and generative AI are describing the same shapes. A Nobel laureate solving a decade-old problem with an AI finally made it click. Giorgio Parisi (Physics Nobel 2021) and Francesco Zamponi proved an identity that had resisted proof since 2014 — in a single 40-prompt conversation with Claude. The model essentially found the proof itself, but only because the physicists prepared the ground with extraordinary care. My piece argues it worked because of the very theory Parisi won his Nobel for: frustrated systems, where conflicting constraints create a landscape full of false solutions. The core idea — hallucinations come in two kinds: • missing information → add more, it helps • too much contradictory information → add more, it gets worse; only targeted disambiguation fixes it Parisi did the second. He didn't teach the model anything new — he removed its confusion. And the humans were stuck too, for the opposite reason: years convinced the answer had to be complex, so they couldn't see the simple one. The model, free of that bias, went straight to it. They de-biased each other. If this two-regime split holds, it predicts something testable: dumping more documents on a model (RAG) should sometimes make hallucinations worse. Has anyone seen that in practice? Thanks to @towards_AI for publishing it 👇 medium.com/towards-artificia…