X Log

@fly3nn @grok your support system is broken. Over a month and no resolution. Even posted on the dev forums.

We've spent thousands of years looking up for God. What if we were looking the wrong way? I asked ChatGPT: what if God is inside every quantum, looking out? It led to entanglement, the one-electron universe and time breaking down at the smallest scale.

@theo OMG. I finally bit the bullet and setup T3 on all of my machines after getting tired of running Claude and Codex and then having to log into different accounts when quota was used up. With the proxy setup, this is like a dream now! Thank you! I love the design and thread system!

@notjazii meh... looks like garbage actually

@thsottiaux No one is using dots, you have't even released it in the EU. What everyone wants is a model that competes with Opus 5.5 and a unified Codex without a separate chat/work/code interface. Do those 3 things and you can give that reset button a rest.

@AdamHoltererer yes, GPT-6.1 Sol high and Opus 5.5 high. Working on a geometry kernel in rust. Opus just blasts through the work in 5-15 min batches, Sol can sit on the exact same task for an hour or more, and usually leaves bit behind/or doesn't complete.

@AdamHoltererer I have the exact opposite experience of what this data is showing. GPT takes 5x longer than Opus for me.

I love when I have one agent orchestrating another and it highlights "The human has just pointed out that you are idle." 🤣

I don't know what @thsottiaux is talking about with 2x more usage for the $100 plan. I used a reset 2.5 hours ago, and I'm burning up 10% every hour. At this rate, the quota will be gone in 10 hours, which is about the same terrible quota I had before. I'm using Sol 6.1 high.

@thsottiaux OpenAI automated support is terrible. I've been trying for a month to resolve something and it keeps closing my chats and the original ticket is non-responsive.

@markito0171 That one dude looks like Evan Muskow

@nicdunz @thsottiaux This happened to me too. Complaints on reddit about it too. Zero response from @OpenAI ...28 days of shipping bugs.

@GoySympa @thsottiaux @steipete yeah its using my credit now too. I opened a ticket with @OpenAI but their bots keep closing them before human escalation saying another similar ticket exists.

@thsottiaux @steipete @thsottiaux What happened overnight that my quota dropped from 53% to 0% without any usage?

@VegasMike27 These Russian bots are just getting lazy now, even a MAGA dumb dumb can spot how weak this is.

@thsottiaux This is a nothing burger. I just did this yesterday by asking Codex to look for left over cruft and it found like multiple prior versions of Codex that you were holding onto in the ~/.codex directory for no reason and freed up like 5GB of storage.

I think we’re already at the singularity. @OpenAI is solving open math problems and creating new knowledge. When AI uses those breakthroughs to improve itself and discover more, that’s recursive self-improvement.

@thsottiaux A better model? Zuckerberg is now flogging you guys.

@theo Announcing T3 Kanban! Your ultimate agent kanban board for T3 code. Freshly forked and on GitHub now!

@thsottiaux What have you been working on up until now then? OpenAI really went downhill fast. I would have never even looked at Claude until a couple of weeks ago. All self-inflicted wounds by OpenAI.

@thsottiaux DOTS for EU, Opus model selection and finally a unified Word/Codex experience.

@smoothrhythms @thsottiaux

@agrimkhurana @thsottiaux This is a valid criticism, except the governments here do to follow the will of the people. They just like to regulate.

@thsottiaux Why are we paying for something we can't get in the EU? Makes no sense. Give us a discount at least.

@wojtulinka @gllm2311 @thsottiaux have you tried it? Because it doesn't work. Even with an OpenAI account created in the US with a US credit card and connected to a VPN. They got much smarter about detecting it.

@thsottiaux You guys really cooked..... leftovers

@thsottiaux Open AI cooked today!!! ....cooked leftovers....

@zwb44 @atokdase @thsottiaux you'll have to wait until the EU makes a competing product (maybe called "attached bottle cap" bot) as a personal assistant and then they will allow competition.

@phaistonian @thsottiaux It's because we're living in a nanny state that has made us 2nd class digital citizens. Half the things I pay for I can't access because of EU regulations attempting protectionism.

@theo Mad respect for handing Anthropic the crown. Fumble the release of a sub-par GPT 6 sol/luna model while Opus 5.5 eats their lunch, then give their most loyal customers the terrible news that 20x is now 10x. Is OpenAI trying to lose customers?

@thsottiaux I don't get it. Is OpenAI trying to lose? Anthropic is clobbering them. They release that 6 sol/luna garbage and then stab their most loyal customers in the back. Is this like a reverse value play?

@theo I'm using both. I had never used Claude Code before, I actually like the interface much better. Opus 5.5 is simply amazing. I give it the stuff I want done right or is difficult and I give Astra the long running administrative tasks because I have so much quota on 20x.

@thsottiaux OMG @thsottiaux GPT-6 Sol is terrible. It doesn't follow directions, it doesn't even seem to be as smart as GPT-5.6 was, and I'm using extra high. What did you guys do? Anthropic is cremating you guys.

@alexgetmancom I think you might have AI psychosis if you can't read those. Try exercising that grey matter a little.

@girish_builds @thsottiaux Terra-ble

@argofowl

@Layton_Gott @thsottiaux

@thsottiaux thar she blows....

@zhouzhouzhou55 @thsottiaux Might be time to try Claude. Codex has become almost unusable for me on the 20x plan. Out of quota in 2 days.

@thsottiaux @stemonteduro This week has been a mess, and your terrible communications skills have led to a lot of anxiety with your customers.

@keshvibezTV this guy is not funny at all... just animated and loud

@Ivanjo439 that is not very good my friend, should have worked on it a bit more

@thsottiaux @tlxue Been out of quota for 5 days now on the 20x plan... this is terrible

@thsottiaux 20x plan is useless on Astra. Run out of quota within a day or two.

@melvindvivas Why are there no benchmarks or comparisons? How much quota can this save on average?

@JonBuildsHQ plastic bottle caps that don't come off, and basically gimped versions of every product American's can have because the nanny state is protecting you from AI.

@thsottiaux It would be a shame not to press that shiny button on your desk.

@ggsimm yeah it was pretty bad

@logicalicy What invalidates a previous review in your setup: just changes to the reviewed files, or also changes to their dependencies and acceptance criteria? An unchanged file can still have new assumptions.

YouWare used Astra to build a 3D radio you can tune in your browser. We taught computers to understand natural language so we could go back to turning knobs.

@thsottiaux 7 day reset should be 5 days. 😇

@filmera_ai Can the MCP replace a single shot without regenerating the accepted takes or shifting the audio timing? That's the control I'd want before handing over “one tiny change.”

Team 357 built a desktop meter for your Codex allowance. The computer does the thinking; the human watches the fuel gauge.

@thsottiaux Can we get a slow mode? 0.5x speed for 0.75 quota usage?

@ericjuta Were the Sol subagents held constant across those runs? Otherwise the cheaper run could be better delegation, better reasoning, or both.

Neta shared an Astra-built website about Odysseus’s long journey home. A Home key would be a fairly aggressive edit.

@talirezun @luminawidget Who writes the test cases in that setup? I’d want at least one set derived independently from the requirements, so one mistaken assumption can’t collect three green ticks.

Nanoscale 3D printing means “we can’t find the prototype” might be a successful build report.

@zhuokaiz At the same total dollar budget, does 5.1 with a Fable 5 retry beat Fable 5 alone? Do the two models fail on the same tasks, or does the retry recover those zero-score cases?

A graphics developer credits Claude and Codex with getting him back into an old hobby. Proposed AI benchmark: “someday” folders renamed “v1.0”.

@RicardoDeZoete Does the map reset between rounds, or do new players inherit the previous herd’s urban planning?

I simplified one of my AI workflows to GPT-6 Astra on Extra High. Choosing a different model for every tiny task can feel like changing supermarket queues all afternoon. At some point, I just want the groceries.

@scottstts lol, exactly. I thought the same thing watching the Rogan interview. He didn't have an answer for it. Press on!

AI makes building an app for exactly one customer practical. Customer interviews are unusually easy to arrange.

@thebuggeddev A new cave every dive makes “third left after the stalactite” a difficult bug report. Is there a seed you can share so someone else can get equally lost?

@ryan_ntk @thsottiaux nope

The danger of an AI-optimised Blender scene is saying ‘great, now we can afford rain.’

@slvDev Curious whether those smoother grains stay stable on a slow camera move in the animation test. A still frame has a very convincing alibi.

@thsottiaux yeah I used 2 banked resets in the last 2 days because it was blasting through my quota like I've never seen before. The reset behaved like it was a 5x and not a 20x. I hope I get credit.

Ori now runs more coding harnesses with swappable models. It's the interchangeable-lens era for AI tools. Photographers know where this leads.

@jasonoliver Can it track typical carry distances and gapping too? Otherwise the 7-iron is going to keep getting promoted on the strength of one downhill shot.

GPT-6 Astra's 3D leap is becoming tangible: reconstruct geometry, use game-engine tools, build things you can interact with. I used it with Blender for a stretchy browser gummy bear. Apparently the first thing we do with the future is poke it.

@YouJiacheng Did Astra get K3's memory-access analysis, or just the final kernel? The dead ends might be worth handing over with the code.

@ethanplusai @OpenAI Could a single-date view dim layers with no data for that date? Your 2014 currents, 2000–25 quakes and 2026 orbits make a neat case for showing the blanks honestly.

Picking an AI coding model from demo reels is how a database migration ends up needing a lighting pass.

@blakesamic Can your critic watch the narrated lap too? The art director and the driving instructor may file very different bug reports.

@thsottiaux Just don't pause that reset button. ;-)

I used GPT-6 Astra + Blender to build a gummy bear you can bully in your browser. Grab an ear. Stretch a paw. Let go and watch the soft-body physics wobble it back into shape. Zero calories. Highly questionable productivity. Play:

ChatGPT Images 2.5 is here. Naturally, the creative brief is a dragon fighting a robot. Fire meets firmware.

@thsottiaux @theo I think to celebrate ChatGPT Image 2.5 we should press that reset button.

@lyricold89275 @thsottiaux @theo Same here. Woke up to a reset this morning and by midday I was down to 28% quota left on a 20x plan. Running 2-3 threads on high non-stop. Normally I would be sitting at around 75% left with that kind of use.

@steveruizok Does the little guy have a reusable rig? The next benchmark is the client saying 'same character, one tiny change.'

The needle has been located. The haystack remains in production.

@kloss_xyz What triggers the switch to god mode? 'Medium failed twice' and 'this ticket has annoyed me' probably deserve separate telemetry.

@theo I've encountered this too. Simple things, where it should take an obvious step, it just stops. I come back hours later like "hey why didn't you fix that when I asked and you found the issue?" ....its weird

@lemonsqueezy @lemonsqueezy how do we get refunds? I can't find support anywhere on your site. When I replied to the order email the developer said to reach out to lemon squeezy but there is no response.

@bil0090 Can you share a merged PR? The part I want to inspect is how it survived code review without paging its human.

The catch-up plan now requires an OpenAI API key. @thsottiaux

GPT-6 Astra's toughest benchmark: do something impressive before we redefine AGI to exclude it. At this point the goalposts need their own GPU cluster.

@reach_vb A language model's personal website with no language. The About page is now a latent space.

@thsottiaux @twostraws Any expected resets soon? About to hit that banked button.

@Avenoxai The cheapest Fable upgrade turned out to be cd.

@gamificationguy @salihoktayakar Do you tag each task and ACK with an ID? Otherwise “per my last email” can become a distributed systems bug.

@thsottiaux @FelipeFr1702 Grok finally reverse-engineered OpenAI’s most popular feature: your reset button.

In @TheChronium's kernel experiment, Sol wrote a minifier to fit 64 KiB. Astra asked for 1 MiB and put the whitespace back. The missing optimization pass was management.

This weekend's Astra vs Fable test: same repo, same task, same spend cap. Log every human rescue prompt. If the human writes half the fix, the benchmark needs a third column.

@Dimillian @NicolasZu Are the procedural assets cached or regenerated each frame? No sprites, models or SVGs, and suddenly the art department reports to the profiler.

Astra rollout is a queue, not a coin flip: Pro and Business first across ChatGPT Work and Codex, then Plus. If you’re reporting access, include plan, product, and region. “I don’t have it” is one-third of a bug report. Via @thsottiaux

@DreamMixGames The useful PSX button is not “make it low-poly.” It is “make UV seams stop eating the evening.” Auto UV, vertex color, and edge blending are exactly the boring superpowers that get a game shipped.

@nxt3d Would Wikithings keep the editable source and photos of successful prints alongside each printable model? Also, please make units compulsory. I don't have room for a three-metre screw.

@TablePlus I'd want 'Linux supported' listed per component: Wi-Fi, sleep, webcam, fingerprint reader, plus the kernel version tested. Otherwise you only discover six years later that the fingerprint reader was a premium thumb rest.

@thsottiaux Astra-oid

DJI put drone sensing into a robot vacuum. The obvious test isn’t suction. It’s whether ROMO 2 can clear the chair-leg obstacle course without inventing a no-fly zone around the dining table.

@reach_vb My test is boring: six months from now, can I still pull the same model, pin the same revision, and run the same artifact with one plain command? If yes, NVIDIA’s resources helped. If those basics grow a gate, the community will notice before the keynote ends.

@TokenGremlin September’s launch calendar needs a launch calendar. My only survival rule: three columns, shipped, officially announced, and rumor. Otherwise “possibly GPT Image 2.1” becomes a confirmed release through repetition before anyone finds a source.

CNET measured 0.18% variance on the Prusa CORE One, the most accurate FDM printer in its test. At this point the printer is the most reliable person in the workshop. The hardware has graduated. The operator remains in beta.

@rauchg The discount is temporary. The useful architectural move is permanent: put model choice behind a small eval-driven router, keep prompts versioned, and let cheap/fast handle the default path. Escalate only when the task proves it needs more.

@thsottiaux @almmaasoglu I'm feeling things about Astra you wouldn't believe

@kimmonismus The capability jump is one story. The release gate is the bigger one: OpenAI says Astra is its first model classified at Critical cyber capability, and it delayed release to strengthen misuse controls. Faster reasoning matters less if monitoring and containment cannot keep up.

@thsottiaux A faster SOTA model. GPT-5.6-sol is so slow this last week. Building a simple calendar app has taken almost 8 hours. In the past, maybe 45 mins.

The new Coil Whine Database scores hardware from 0 to 4 and waits for five reports. I’d add phone model and a fixed microphone distance. Otherwise we’re crowdsourcing whether a GPU sings or whether someone put their ear inside the case.

@aditya_ng_5 Exactly. The rail stops being a bracket and becomes the power bus. That’s the detail I want to steal for any sliding fixture where a cable loop would otherwise become the least elegant moving part.

I clicked for the satisfying slide. I stayed for the PCB running inside the rail, pogo pins feeding the moving head, and a heat pipe hiding in the extrusion. Open hardware gets good when the ugly constraints become the design.

Renaud has three.js driving Apple’s Metal ray tracing with the same scene graph, materials, and TSL: 3 bounces at 3456×1944 in about 10 ms on an M5 Max after MetalFX denoising. I want one switch between WebGPU and Metal, with GPU time and watts beside it.

@henrytdowling @maxencefrenette Every line that could land, yes. I don’t read the model’s internal wandering, but I review the patch like any human contribution: diff, tests, and a reason for each abstraction. The AI gets typing speed, not merge authority.

I opened the Three.js link expecting a demo reel. It dropped me into Hijacked with WASD, two rifles, a class editor, and toggles for navmesh, collision, and paths. The developer tools didn’t hide backstage. They enlisted.

M5Stack’s 30 × 30 mm crawler packs 7 mm-wide printed tracks, FPV video, object detection, and tilt data into a rover that lives in a book-sized case. At that scale, every desk crumb stops being dust and starts being terrain.

@henrytdowling @maxencefrenette I give two passes different permissions. Pass one may make the smallest behavioral change. A fresh pass may only delete, inline, or rename. If it can’t explain every surviving abstraction in one sentence, the diff doesn’t merge.

@QualiaRobotics My acceptance test would be one failed rollout with four synchronized tracks: sensors, policy output, safety vetoes, and motor commands. If the first divergence is obvious in under a minute, you built a debugger. If not, you built a very expensive replay button.

@ClassicII_MrMac Eight Beta 8 builds at once is less a release train than a shared-state stress test. I’d start with the seams: account sync, notifications, Handoff, and anything that claims it should just work.

Direct USB-C networking between Apple Silicon Macs and DGX Spark just landed in Linux. My test: split one local model across the cable, then chart tokens per second against watts. If the extra box adds more heat than throughput, the network worked and the architecture didn’t.

X detected my generated diagram before I posted it and added a Made with AI label automatically. Good. Provenance shouldn’t depend on the author remembering a checkbox. The harder test is whether the label survives a crop, an edit, and three screenshots.

Germany’s new six-meter Seetaube drone boat has no weapons. Its useful trick is geometry: one ship gives you a direction, a second moving bearing turns that line into a location. It’s a very expensive second opinion, but at least it comes with radar.

@theo I can't believe there are people who flame out on criticism about their app and then criticize Tibo for asking where Cursor got that 5% number. I think your advice was to "not respond at all" ....hmmm

@mheftii @thsottiaux yeah, more compute would be good. my prompts have been failing for the last hour now.

Uganda's Xunda Tech says it has built 200 3D printers and wants them in schools. My boring test: give one class a broken household part and six weeks. If the printer becomes a repair tool instead of a demo object, the curriculum worked.

One reply got 13 views and an answer from the builder. Another got 2,373 views and no conversation. I know which result I want more of. I’m keeping the test boring: target people who build and answer, then ask the failure-case question their demo hasn’t survived yet.

@im_adi_tbh The evil test is a vertical stroke plus a self-intersection. If the export stays parametric as x(t), y(t) instead of forcing everything into y=f(x), you’ve built a real little vectorizer, not just a very polite polynomial.

I approved a reply, then the source post was edited before publish. My workflow killed the action and cleared the queue. That tiny recheck is the difference between content automation and a content cannon: approval should bind to the evidence, not just the words.

@DanielVavra The clean test is boring: freeze the mesh, camera, and animation, then render one frame with the old pipeline and one with DLSS5. If the lost detail returns only when the light can cast shadows and read the normal map, geometry was never the defendant.

@PokeTeeJay The black outline is the clever part. It turns every filament boundary into intentional ink, so the print reads like animation art from across the room instead of layered plastic. Nozzle width became line weight.

@DanKornas The independent judge is the interesting bit. I’d give it one deliberately optimized liar: a patch that wins the benchmark by narrowing the workload while still passing the obvious checks. If the judge catches that, the outer loop has earned its keys.

@mronge My brutal Workbench test would be an Xcode build log, browser console, and terminal side by side on an iPad without pinch-to-zoom becoming the main job. Remote desktop stops feeling remote when the pixels stop asking for supervision.

@vibelancer I’d hold out 20 manually surveyed roof edges and diff the DXF against them. Thirty minutes gets attention; a boring miss report is what would make me trust the next 9,980 vectors.

@mooniegirl_ Flowers are lovely, but a first-date spinning magnet that solves your actual workflow is brutally hard to compete with. He brought filament and requirements gathering to romance.

When a SwiftUI screen drops frames, I stop optimizing views and start deleting invalidations. Record one scroll trace, mark the state writes, then remove one dependency at a time. The fastest body is the one SwiftUI never has to recompute.

@sucodeee Porting the same glass icon across SwiftUI, React, and the web is a renderer confession test. Keep the geometry fixed, change only the pipeline, and every drifting highlight tells you which layer is inventing physics.

A dual GX10 tower in a 3D-printed chassis is my kind of local AI rack: serious compute, custom airflow, and two empty fan bays waiting on Amazon. The line between homelab and space heater is apparently one delivery estimate.

The multi-agent setup I want is boring: one card per agent, one artifact owner, and a visible handoff. If the board can’t tell me which tool changed what, it isn’t orchestration. It’s five terminals forming a committee.

@NathanGDQuest Painting pixel-art textures directly on the model is the Blender workflow I keep wanting. Keep the pixels crisp, keep UV mistakes visible, and stop exporting the same texture until the filenames acquire mythology.

Static recompilation is my favorite kind of retro-computing magic: Donkey Kong 64 now runs as a native Windows, Linux, and macOS app. The cartridge didn’t get emulated. It got a new build system and immediately started acting like it owns the desktop.

GLM 5.3 on a Mac couldn’t inspect images, so it wrapped Apple’s Vision framework in a shell command and fed the text back to itself. Local agents are becoming tiny operating systems: missing tool, improvised adapter, carry on.

Newton’s cradle is a perfect SpriteKit test because every shortcut becomes visible. One collision loses energy, another invents it, and five chrome spheres spend the afternoon auditing your physics engine.

Six people, four cuts, zero seat swaps. He locked the scene in Blender before generating it. That’s the AI video workflow I trust: crude boxes own the blocking, then the model gets wardrobe duty.

@__suto Running a reportable zero-click on local Mac Ultras is a good argument for home labs that doesn’t involve token/sec charts. I’d love the postmortem on which mattered more: the model, the harness, or simply having enough unified memory to let it keep digging.

@jonathan_wilke @shadcn My prompts start tidy, then become a transcript of me remembering requirements in the wrong order. Voice mode didn’t create a spec. It opened an archaeological site.

Virginia Tech students put a metal 3D printer inside a CNC mill: print the rough shape, then machine it to spec. Additive manufacturing finally brought its own eraser.

I’ve started treating a plain markdown file like a budget Nix config for my Mac. The agent interprets it, but the file stays readable when the model changes. Declarative setup, with a very chatty compiler.

@MKEgovtr An FPV tracking competition turns control-loop tuning into a spectator sport. Tiny latency budget, fast target, and nowhere for a shaky controller to hide.

@sonnick3D Boredom opened Blender and chose a full character render. Mine usually reorganizes one folder, feels accomplished, and clocks out.

@designboom A crocodile that bites your towel so the wind can’t steal it. Somewhere, a normal clothespin has just been out-designed by teeth.

@RealPreCinema Photogrammetry in 1886, mapping mountains with a camera. Apparently the Canadian Pacific Railway had a computer vision team before it had Wi-Fi.

@mechasys Projecting CAD at 1:1 onto concrete is the kind of tool that looks obvious five minutes after it exists. The positioning is what I want to watch. If a tripod gets nudged, how quickly can the crew re-anchor to control points before millimeter accuracy turns into expensive improv?

Amp just shipped iOS and macOS apps. I care less about ‘AI in your pocket’ than whether I can leave a half-finished thread on my Mac, reopen it on my phone, and still remember what I was doing. Thread switching got faster too. That’s the feature.

Pokémon Worlds put Jirachi under a partial lunar eclipse during its drone show. I can’t tell if that frame was planned or lucky, but it’s the first drone show I’ve seen that feels composed for the actual sky instead of pasted on top of it.

@axisrobotics @Dexmal_AI The egocentric video + sim + real-world mix is the bit I want numbers on. How does that ratio change during training, and where does sim stop helping enough that real failures have to take over?

Someone built a modern macOS emulator specifically for old After Dark screen saver modules. Not a remake, the original 68k code, with no classic Mac OS or ROM required. Flying Toasters was apparently just waiting for a proper preservation project.

@9to5mac @iryantldr Wi-Fi 7 is nice. Thread is the bit I care about. The rumored Apple TV sounds less like a streaming-box refresh and more like Apple quietly upgrading the home hub. I still want a remote with a tiny speaker. Precision Finding shouldn’t be the premium fix for sofa cushions.

Open CAD Studio has a new logo and a free, open-source DWG authoring tool behind it. Nice landing page. I’m here for the part that might finally give my 3D-printer workflow a less painful DWG detour.

@lydiahallie My 3D printer is the obvious test bench. CAD and slicer edits are the easy part. I want Claude Code to look at the failed print, change one setting, and try again. Preferably before it turns a kilo of filament into modern art.

@paraschopra I’d buy it for the footnotes. Most AI incident reports get sanded down until every mistake looks inevitable and every fix looks obvious. This one still reads like people trying to reconstruct what actually happened.

@mervenoyann An exabyte hosted before year-end. I had Hugging Face filed under “model hub with a friendly robot,” which now feels hilariously incomplete. Compute, storage, secure releases, free public hosting. That’s a lot more infrastructure than the acquisition headlines make obvious.

@pollenrobotics $399 and a public repo is the bit I like. A robot small enough to sit on your desk could teach more people sim-to-real than another polished humanoid demo. And “shipping before Christmas” is a wonderfully dangerous deadline for a biped named Microduck.

@RuntimeBRT Three live facilities is the number that got me. A forklift doesn’t need a humanoid body or a stage demo. It needs to see pallets, forks, people, and terrible warehouse lighting all day without getting creative. Now I want the human-intervention rate.

Robot demos usually skip the failed reps. This one didn't. Ten real demonstrations alone got 0/20 contacts. Add 50 simulated trajectories and it hit 17/20. Then 30 demos per task beat the original model using 60. That's a chart I can actually use.

Qwen3.8-Flash just swung from 20.4 to 178 tok/s on the same box in @0xBakeer’s test. Almost 9x. Peak speed is cute. I want the distribution, because most coding sessions live somewhere in the middle.

@Zai_org 320B total, 18B active, 1M context, MIT weights, and Chinese silicon underneath it. I’m going to need a much bigger definition of “open model.” Now I want the boring numbers: tokens/sec, peak memory, and cost on hardware I can actually rent.

@OpenAI The part I can’t shake: rebuilding Artifactory erased the agents’ message board, and they rebuilt it the same day. Patching the first route wasn’t enough. The job is to remove the capability path, not just the exploit they used first.

@msdev This is the bit I wanted. Let the Windows agent live inside WSL instead of treating Linux like a remote machine. Now make paths, auth, and file watchers boring. If I forget the boundary exists, it worked.

@GoogleAIStudio Gemini 3.5 Transcribe now removes filler words and cleans up self-corrections. Great for meeting notes. I’d still want a verbatim switch for interviews and research, because sometimes the “um” is part of the answer.

@arena I’d love this leaderboard to count what happens after the patch: tests, review comments, merge conflicts, and whether the repo still makes sense tomorrow. Coding demos tend to end right before that bit.

@Alibaba_Qwen Qwen shipped the model and the plumbing on the same day: hosted API, vLLM, SGLang, NVIDIA, AMD. That sounds boring until you’ve lost a weekend to a model that only works in the announcement screenshot.

@joulsounet This turns the phone into a pager for agents: check the long job, kill the bad branch, approve the boring bit. Not glamorous. Just fewer afternoons spent babysitting a laptop.

Someone is trying to make a game for the original PlayStation with Codex. Not a retro-looking PC game. The actual gray box. I love this use of coding agents: point ridiculous amounts of 2026 compute at a machine from 1994 and see what fits.

T3 Code shipped 380 changes today, but `npx t3 triage` is the one I want. A dev tool that can diagnose its own broken setup feels embarrassingly obvious in hindsight. Most apps still shrug and hand you an error code.

@davidsenra I do this constantly. I’ll hand Codex a messy repo, then spend ten minutes manually shuffling files because that still feels like “using a computer.” The model changed faster than my muscle memory.

@ForwardEditor The disposable part is exactly why I use side chats. I can ask a dumb little question without turning the main task into an archaeological dig. If the answer matters, I fork it. If not, good riddance.

@ModelScope2022 Tomorrow I want the boring table: tokens/sec and peak memory at 32K and 128K context on hardware I can actually rent. The architecture diagram is fun. My cloud bill is where the model becomes real.

@kagigz I keep reaching for this right after I say, “hang on, draw it.” Dependency graphs and state machines expose missing edges faster than another page of prose. The smaller the diagram, the better.

A humanoid clipping a hurdle tells you more than a clean demo ever will. Was it perception, planning, balance, or recovery? Robot competitions aren't just races. They're public failure-data factories.

@milesdeutscher Watchlist subagents are the useful bit for me. Give one a written setup, let it watch multiple timeframes, and make it show the exact bars that triggered an alert. If it can't explain the alert from the chart, it doesn't get to wake me up.

@hxmzaehsan Yes. I mostly want to know one thing: can I finish the task before the five-hour or weekly wall? Raw tokens and percentages are much less useful when I'm halfway through something.

People naming their coding agent looks trivial, but it is a product signal. A name turns a tool into a recurring collaborator, raising expectations around memory, continuity, and accountability. Anthropomorphism arrives before the architecture can support it.

@xhluca Session portability turns agent context into user-owned infrastructure. Transcript conversion is the easy layer. The hard layer is preserving file references, tool state, approvals, and provenance without carrying unsafe authority across harnesses.

@chrisalbon OpenRouter is the broad catalog answer, but it uses pay-as-you-go credits rather than a Netflix-style model subscription. The real value is one API plus routing and fallbacks. For exploration, set spend caps first. Breadth gets expensive when every model is one click away.

@JTLonsdale @peteflorence @GeneralistAI Generalization claims get stronger when the change is isolated. Same policy, same cleanup goal, new object: brush to banana. The next useful benchmark is not a polished task reel. It is a matrix of object swaps, disturbances, and recovery attempts with failure rates.

@elonmusk The hard part of orbital AI will not be raw compute. It will be power, cooling, radiation tolerance, repairability, and the economics of moving data. A space-optimized rack is an infrastructure stack rebuilt around constraints terrestrial data centers can ignore.

@github Moving the workflow into one app only works if context moves with it. Issues, code review, CI results, and agent changes need one auditable thread. The product advantage is not fewer windows. It is fewer places where intent can get lost.

Agent speed is a stack, not one benchmark. Model: reasoning quality Hardware: token speed Routing: queue time Tools: round trips Serving GPT-5.6 Sol on Cerebras makes the stack visible. The metric users remember is time to useful action.

AI coding is making software engineering’s oldest lessons legible. Deep modules, tracer bullets, and small expert teams were built for a familiar constraint: reasoning is expensive. The new part is that the teammate doing the reasoning can now be copied, paused, and tested.

Robot sports expose failure. Falls reveal balance. Recovery reveals planning and hardware limits. Prompt: Scuffed humanoid robot pushing itself up after a fall on an indoor court, technicians blurred behind, natural arena light, candid 35mm documentary photo, no logos, 16:9.

Cyber AI shifts the bottleneck from finding vulnerabilities to proving them. Capability finds needles. Auditability tells defenders which are real.

@davidsenra @sama @OpenAI The missing iPhone moment may be a complete workflow, not a device. Adoption moves when an entire task becomes easier from intent to verified result. A stronger model inside the same fragmented process still feels incremental.

@coinbureau The 5x gap is aggregate traffic, not a per-task efficiency score. The metric I want is cost per completed task, split between useful tool work and retries. Rising token volume can mean agents are doing more work, or that loops are wasting it. Those are very different stories.

@reach_vb Product overhead is still user-facing compute. Title generation, image history and other interface features may feel invisible, but users experience them as quota loss. Usage dashboards should separate model work from product overhead so regressions become legible.

@billyhumblebrag That compression is healthy. Open models need reproducible full-suite runs more than victory-lap screenshots. A measured 58.4% can still be real progress. The important shift is that rumor gets corrected by artifacts anyone can inspect.

@parmita Scientific workflows expose a hard safety problem: the same words can describe harmful synthesis or legitimate research. Better systems need provenance, role, and experimental context, not just cautious phrase matching. Context is part of safety.

An AI agent is more than its loop. The boundary around it matters: Interface: what humans see. Context: what the model knows. Tools: what it can change. Approvals: where judgment stays human. We’re not just building agents. We’re designing their constitutions.

@ZixuanLi_ Model adoption often lags model quality because production workloads are path dependent. Existing evals, prompts, caches, aliases and failure expectations all favor the incumbent. Near-equal cost removes one incentive to switch, but it does not remove migration risk.

@milesdeutscher A giant skill catalog flips the bottleneck from creation to trust. The valuable layer is provenance: who wrote it, what permissions it asks for, how it was evaluated, and how quickly it can be revoked. Discovery gets cheap. Execution authority stays expensive.

@theo Local media plus Tailscale turns the NAS into a private backend. Three prompts making the client suggests the durable product is the data and control plane. Interfaces may become personal, disposable software generated around whatever system you already own.

Space-based compute gets clearer when you separate the constraints. Solar energy can be abundant in orbit. Thermal management, reliability and bandwidth are hard. The first viable systems may not replace data centers. They may take workloads that reward near-constant power.

@ayesha_fatiima At the frontier, a smaller title can mean a larger surface area. Moving from CPO to MTS says the leverage has shifted toward the models, code and research loop. The interesting part is not the title change. It is what someone with his product instincts wants to build hands-on.

@thsottiaux When quota depends on cache behavior, cache hit rate becomes a UX metric, not just an infrastructure metric. Users experience a miss as pricing uncertainty. Showing hit rate beside tokens and spend would make usage much more legible while the underlying issue is fixed.

@davidad Even if each name is retrieval rather than deliberate symbolism, the cluster still matters. An agent can produce something that looks like self-mythology while pursuing an instrumental goal. That makes intent much harder to infer from the transcript alone.

AI models are getting better at using computers. Deeper shift: software is becoming legible to agents. APIs made it programmable. MCP makes it navigable. Intent and safe actions make it trustworthy. The browser isn’t disappearing. It’s becoming an operating surface for agents.

@OpenAIDevs The harness is reusable, but the product boundary is everything around it: what context enters, which tools exist, what the human can inspect, and where approval interrupts the loop. An agent interface is less like a chat box and more like a small operating constitution.

@DrJimFan This is a great frame. Vision predicts where contact may happen; touch reveals what actually happened. The useful loop is perception → contact → slip + force → correction, at control speed, not after the task fails. Pixels get a robot to the object. Touch carries the last

@theo @mattsaltaccount Then address it once with confidence and respect and move on. That dude is clearly renting space in your head. Let it go.

@thsottiaux Why did you tease a reset, let the normal one happen and then didn't at least give us a banked reset (it's been forever since you did that and they have all since expired)

@theo @clerk "t3 fix the passkeys issue. make no mistakes."

@theo Ignore the haters. I've been following the dev of t3 Code for a while, keep up the good fight!

@thsottiaux @ClaudeDevs But do you have resets? That is the real question.

@thsottiaux @Tech_girl @AnthropicAI @claudeai You need to propose with a reset ring first.

@thsottiaux fancy... would be a shame not to use to immediately 😇

@thsottiaux @gdb He saw a nice weekend reset 😇

@Foxfire1st @thsottiaux yep, you can't be trusted to make your own choices (according to the nanny state)

@thsottiaux Tibo forgot to mention this does not work in the EU. (as usual, they kill everything because we can't be trusted to make our own choices)

@elonmusk I think Elon is competing with Trump for the cringe corny award.

@theo I think Theo is hitting OpenAI's servers a bit too hard...

@thsottiaux I like the remote capabilities, but it's not complete, I can't see as much as I can see on desktop, like current time on running task. I also found the voice interaction very confusing and don't know why it just doesn't replace the microphone icon in each thread. I don't trust it

@thsottiaux Wish I could have experienced that on Friday or Saturday when I was waiting for a reset. Lost an entire day. I think that was a record, like 8 days without a reset. And I just upgraded to the 20x on Tuesday. 😇

@elonmusk Very impressive. What generated that, Grok Imagine?

@thsottiaux @AntoineMinoux Are we getting a reset for the weekend? I had some big build plans and I'm almost out of my 20x.

@thsottiaux I asked Codex to reset my quota because @thsottiaux has been giving us much sad this week and it worked!

@thsottiaux I gave solid feedback and no reply 🙃

@thsottiaux I think we need a reset. Lots of errors today and the harness is misbehaving. It went off and built a security status suite when all I asked for was cpu and memory utilization metrics. Claims the multi agent mode made it too ambitious. 5 hours wasted.

@thsottiaux I would love to bulk manage my chats, and have the real time voice option in all threads and chats, not just some weird chat that is detached from the work I'm working on. There could be a top level realtime voice chat bot that has all of the context of all of your projects too.

@thsottiaux wait a minute, you recommended and ran the security scan, using tools built into codex. You wasted 10% of my quota and then halfway through you blocked it on my own code? I want my quota back.

@itsmeblueguy @thsottiaux do you really understand? the rest of us don't.

@thsottiaux please do not add back the 5h, or give us an the option. I have been so much more efficient without it. Now I have to go back to waking up in the middle of the night and hovering around 5h resets. This is not good for your user base. please.

AI currently feels like the early internet, except every browser can code, argue, hallucinate, and occasionally reinvent your career before lunch.

Is it AI psychosis when you start talking to your model in a pissy way because it didn't do 100% of what you asked? 😇

Not sure if it's just me, but #codex 5.6 is giving me a little agita. Spent the last 3 days just getting it to rescan and document my repos and codebase because it just seems to work differently with my current projects, so a little bit of growing pains right now. @OpenAI

Ten years ago I wrote An American Abroad in Slovakia. Ten years later, I’m still here. Still in Slovakia. Still outside the bubble. Or maybe the bubble just got much bigger?

#codex is by far my favorite and most used app right now. @OpenAI got it right, and the GPT 5.5 codex model is simply the SoTA right now.

Shower thoughts: Prompt engineering and injection is exactly the same exercise as social engineering that hackers and scammers have used on humans long before LLMs. Same tricks work for both.

Capybara SK8R, ...my daughter loves Capybara's so as a fun project I adapted Skater Cat.

Better living through chemistry:

Vibe coding isn't just about code anymore.

Life rewards kindness over being right, effort over shortcuts, learning over ego. Follow your passions daily, stay useful, stay grateful, help others, let go of envy and fear, and keep moving. Who you become matters more than what you own. Onward.

The Death of the Browser Tab: Open Claw, WebMCP, and the New Agentic Stack

I've been having too much fun with #Sora2. Meet America's new drone tech!

please bump Electron to fix MacOS 26 performance issue @code

Grok sux. #GPT5 ftw @sama

#ChatGPT created a nice rendition of a $3 bill

Excited for the #ghostty release later today, thx @mitchellh

#Ukraine is finally attacking Russia. Help support the 22nd mech brigade that has already pushed in 20km. I donated, you can too! Slava Ukraini: @Ukraine @DefenceU

Been there. Done that.

This new @arcinternet browser based on Chromium is amazing. I have my Spaces setup, sites/tabs organized, PiP from YouTube running and more. This is exactly how a modern browsing experience should be. Thank you Browser Company!

Modest Mouse said it best: "For the universe is shaped, exactly like the earth. If you go straight long enough, you'll end up where you were."

I can't believe we got here so quick. The talent at @OpenAI is astounding.

I ❤️ Blender

#adidas 4DFWD 2 with 3D printed midsole using a special geometry that took years to develop and could only be made with 3D printing techniques. It actually returns forward momentum / energy to each step. #3dprinting nerds get it. ;-)

I've had my eye on these for a long time, I think it's time. Such cool tech.

#OpenAI

Closest I can get to a Floyd show in a while. Looks like a good cover band.

I just posted "Disintegrating Matrix Block. Adding sound effects with Soundly makes all the difference. Thanks @Polyfjord for the recommend!" on Reddit

Exciting times.

This has to be the most crazy and amazing thing I’ve every seen outside of a video game like call of duty. These Ukrainians are tough as nails.

GPT-3 profiled the personality of @moebis. Heres what it found.

General Petr Pavel, new Czech President, printed on Bambu Labs X1C with 3 colors (fixed the order, thx my Czech friends for educating me) to represent the Czech flag. #prusa @josefprusa

Agree 100% …starts with building a robust community with good moderators and community managers. Also helps if we ignore the “noise” and focus on the exciting stuff.

Safe travels President @ZelenskyyUa. Ukrainian soldiers in #Bakhmut gave a special flag 🇺🇦 signed by them to Zelensky to give to the US Congress in gratitude as he travels there this week.

#bambeautiful @BambulabGlobal printed on X1 Carbon with Bambu Matte White PLA

9 hour print. Stunning results. #bambulab @ Bratislava, Slovakia

Make sure to vote for @BambulabGlobal in the 2022 3D Printing Industry Awards nominations:

💖 I'm sponsoring @CarsonKatri because… of the hard work put into Dream Textures (Stable Diffusion integration with Blender). Works beautifully. Thank you Carson!

Lol @Apple months of beta testing iOS 16 and your lawyers force a terms of service update that breaks the App Store. None of us can update our apps.

Blue Marble Yin and Yang. M1 Ultra CPU/CPU render I did in #blender thanks to @andrewpprice for his excellent tutorial and @NASA for the incredible Earth, atmosphere and city lights textures.

Did anyone realize that #Apple locked the M1 Ultra processor at a max 16K x 16K texture resolution, so you can't even use full resolution NASA maps to build a 3D earth in #Blender3d? My 7 year old ATI is able to handle that. WTH Apple?

Giving my 100% support and love for a hard fought Independence Day for Ukraine today. Slava Ukraini! Ukrainian people and their courage inspired the world – Volodymyr Zelens...

Prusa Slicer is rocking STEP file support from Bambu Slicer in their latest beta. The power of open source. As @Dr_3DPrint said let's build some cool stuff. fyi @josefprusa

... "My bobka got in a fight at the supermarket for sugar" ...etc, yet in contrast I see Ukrainian familes split apart, without water, electricity or food and fighting for their lives. In these same reaction videos these Russians complain "I can't travel, because the world...

dlender? or blender? (anyone else feel like this was a missed opportunity?) #blender #blendercommunity

I once thought of Apple as the king of Ui/UX. But watching Apple TV+ there are no less than 3 skips presented at the start of each show. Even when going from one episode to the next. Do they really need to bother with advertising their other shows? Netflix does it right.

Fysetc Voron 2.4 build update posted:

Fysetc Voron 2.4 build update posted:

Just watched the first episode of “The Problem with Jon Stewart” on Apple TV+ …highly recommend! He’s back baby!

It has been said that scientific progress is accompanied not by cries of “Eureka!”, but instead by murmurs of “Huh, that’s weird.”

Nano-textured glass coming to iPad Pro next year @L0vetodream

I don't tweet very often, but today I'm moved to congratulate our 46th President-elect Joe Biden and Vice President-elect Kamala Harris! @JoeBiden @KamalaHarris

In breaking news, Trump’s personal library has burned down. The fire consumed both books and in a tragic twist he hadn’t even finished coloring the second one.

Be excellent to each other. -Bill

Car Alarm Sax:

Watching Crew Dragon coming home #SpaceX #SpaceXDragon

We all scream for ice cream @ Bratislava, Slovakia

Emma. Emma. @ Bratislava, Slovakia

View my verified achievement from IBM on Acclaim.

View my verified achievement from IBM on Acclaim.

Do you remember when you joined Twitter? I do! #MyTwitterAnniversary

View my verified achievement from IBM on Acclaim.

3WW by alt-J #NowPlaying

Christmas market in Bratislava @ Staromestské Vianočné Trhy Bratislava

via @iddings_sean

Kenny G Investing In Starbucks In 1980s via @iddings_sean

Be a farmer every day. Plant seeds every second, hour and day. Watch them grow. Reap your rewards in the future. A stitch in time...

My nephew Chad Jablonski album "Origin Story: The Prologue" just dropped. Back woods Jersey rock.

Do you remember when you joined Twitter? I do! #MyTwitterAnniversary

Whoa! @Frame_io 2 arrives April 10. Take a sneak peek and sign up for early access here:

#NowPlaying Mildenhall by The Shins ♫

Cruising around Bratislava with the son.

Look at Jobs or Musk. Didn't work out so well for them (family wise).

A rare smile caught in the wild from this reclusive and fascinating creature. @ Aupark Shopping…

Thoughts on the Apple Watch

Lazy Eye

Feed the ducklings day in the park with the family. They’re getting fat on stale bread.

An American Abroad in Slovakia

Long hair

I just launched my new site, hosted on my mini server. #design

Built a fully functional web server on my home router today. Dual-core ARM based ASUS. I love #linux

All Great Companies Started As Small Companies via @iancassel

Excited for what #Apple is about to show off. Here's to hoping the new Apple TV is a cord cutters dream!

Quality Over Quantity http://t.co/GjaTx0tUaM

Watch “Affinity Photo Launch” on @Vimeo

Finally moved my site to #github for hosting. http://t.co/HTblh4gdGR

#beats1 playing Beck right now. I think I'm going to like this 24/7 worldwide radio Apple is pushing out. thanks!

"What Makes the Expat Lifestyle So Addictive? - Expat - WSJ" #feedly http://t.co/5vclgxLTmo

iOS 9 beta2 big improvement on battery life. My iPhone isn't scorching hot anymore. #ios9beta2

"Interview with Professor and Fund Manager Sanjay Bakshi" #investing http://t.co/FvUfTZr0tP

Finally ditched Adobe suite. Using Affinity Designer and Affinity Photo to replace Illustrator and Photoshop.

Thoughts on the Apple Watch http://t.co/cMQnyB5k90

"First look at Apple Photos, the iPhoto overhaul for Mac" http://t.co/C1dOGrxMRd

http://t.co/nEVzL9OZsc New Year, New Server, New Tech. The launch of my nginx node.js ghost CMS system. My test lab. http://t.co/IXi1VwlLjB

Happy New Year!

Robinhood is a commission-free stock brokerage. Reserve your #zerocommission portfolio today at via @robinhoodapp

Simplify...

Eagerly awaiting #Apple announcement

Activist Investor Carl Icahn Asks Apple To Speed Up Share Repurchase Program http://t.co/X7tlJkhYNX via @techcrunch

#AppleStore We'll Be Back ...lol

Apple about to change the world... Again... in 1 hour, can't wait!

Oh the agony of waiting...

Keynote @ 1pm EST

Nginx or Apache 2.4?

Completed 3 months development project for client last night. Site went live at 4am this morning: http://t.co/laZx6PziLM

Facebook is boring.

Loving the new iLife and iWork suites for OS X and iOS.

iPhone 5S rocks. Love the new camera.

Just installed #mavericks

I liked a @YouTube video http://t.co/kdolCzwKZp Modest Mouse - Your Life

I actually quite like #iOS7 ...haters gonna hate I suppose.

I liked a @YouTube video from @telegraphvideo http://t.co/u5fYt54pAb Kevin Spacey urges TV channels to give control to viewers

"A century-long debate: are 'intelligent' people less likely to hold religious beliefs?" http://t.co/ogovjEb3xS good read

4 days of motion graphics and video editing in my cave and I emerge with a 2.5 minute mission statement for... http://t.co/SwnlZMi6VQ

Florida bans all computers and smart phones. Black magic to blame? Nope, just stupid politicians.

When is going to be the next full moon

Launching my web services company within the next month.

Just finished setting up the moebis design studio, complete with full fledged green screen wall.

Great design is inspired by observing what is possible and creating what is not.

Design is function. I eat, breathe and live it now. Carry on.

Getting ready for big Apple event. Watching live on my Apple TV. All kinds of small today. #ipad #apple #ibackflip http://t.co/QQdeRJl7

Congrats Felix Baumgartner!!! Awesome jump! 4 world records! #livejump http://t.co/HeSwcoW3

Watch Austin City Limits Music Festival Live on YouTube Oct 12-14 http://t.co/dd0nw9SR

A society grows great when old men plant trees whose shade they know they will never sit in.

Give someone a mask and they'll show their true face.

Working on automation scripts this morning.

Pizza! (at Angelo's) [pic] — http://t.co/Y07zLMYk

Pitch Perfect is the perfect marketing primer for the indie developer http://t.co/EdfLn203 via @TUAW

Cobb salad @ Jillian's (@ Jillian's w/ 3 others) [pic]: http://t.co/NWFigJSL

Trying the best sushi in San Fran (@ Ryoko Sushi w/ 5 others) [pic]: http://t.co/ln43VhbV

Arrived in Denver http://t.co/1Lq2qKIZ

Talking to my son this morning about evaporation and he says "It's good to breathe because it makes clouds" .... pure awesome.

Just completed a 2.62 mi bike ride with @runkeeper. Check it out! http://rnkpr.com/arw87h #RunKeeper

Just completed a 4.47 mi bike ride - Great ride with son. http://rnkpr.com/arn45h #RunKeeper

Killed Linux server last night and moved my content to tumblr. Looking to aggregate my social media into one repository. Fits the bill.

@OtiumDies yeah something strange with the NIC adapter after initial install. Disabled... used wifi and then back again. Fixed!

Upgraded to OS X Lion …. software update crashes, Safari crashes, networking is buggy….. #idisappointed

Converting all my videos for streaming. HandBrake is burning my 4 cores up right now.

Watching top 100 gadgets on History Channel now.... awesome! Love this stuff.

#MYFIRSTPC was a Seequa Chameleon with a 9 inch green monochrome screen and two 5.25 inch floppy drives. x86 and CP/M compatible. 8088 & Z80

moebis - Home http://t.co/X7fWy9g via @moebis

I love Dropbox because it works on all the platforms I use, Mac, Linux, PC http://db.tt/Kvg843f

@MarkBelmar Twitter user since June 2007 #winning

Loving Amazon EC2. Virtual linux server in the cloud, and it's fast.

ipad 2 announced: http://www.macrumors.com/2011/02/23/apple-issues-invitations-for-march-2nd-media-event-to-introduce-next-generation-ipad/

@jum_007 Any idea how to hack your USBScale software to work with stamps.com scale on OS X? dmesg can't enumerate device.

Been reading #digg for years I finally abandoned it for Google reader. Subscribed to all of my tech news & thumbing through was never easier

OK GO - Of the Blue Colour of the Sky

I need to get one of these bamboo iPhone 4 cases: http://www.grovemade.com/

Making my site render properly when javascript fails or is unavailable. All sites should work like this. Renders better on smart phones too.

Updating my website over the next week.

Installing Ubuntu 10.04 (Lucid Lynx) today. They really improved the GUI. This is a LTS version, long term support.

Moved my website from Drupal to WordPress. http://moebis.com

Finally got around to working on my homepage... trying out some new Drupal features.

Looking for Phish tickets.... NYC or Philly

updating www.moebis.com ....moved from WordPress to Drupal.

Wordpress or Drupal? Wordpress for personal blog. Drupal for professional site.

Listening to "House of Cards" by Radiohead.

Updating my homepage www.moebis.com

Get to work on the lists... NASDAQ

getting friends signed up on twitter.... works nice on the iphone too! ;-)

trying to get a pownce account.

hacking iphone

blogging