silv.blog

Flat illustration of a conference stage at night seen from the empty back rows, a single warm spotlight on a lectern where a fluffy yellow round creature in round glasses and a black bow tie gives a talk, with a few small glowing dots drifting above the stage

AI weekly: OpenAI dots, America.gov, Claude Opus 5.5 and Sonnet 5.5

358 tweets liked between September 20 and September 29, filtered down to what’s worth knowing (here’s how the pipeline works). AI for Everyone OpenAI Launched Dots at DevDay, and My Feed Shrugged OpenAI introduced dots, always-on agents powered by GPT-6 Astra, its top model. Each dot has its own cloud computer and browser, works across 4,000+ apps through ChatGPT plugins, and runs around the clock, and you set what it does alone, when it asks first, and what it never does. OpenAI lists Pro, Business Premium, and Enterprise as the plans, but as of 5:45 pm Eastern on Sep 29 nobody I could find had access yet. OpenAI’s Tibo Sottiaux says it won’t draw down your usage limits. Sam Altman had teased “a new thing” the night before, and the reactions I liked afterward were mostly flat. Ibrahim Ahmed called DevDay “super underwhelming” and short on ambition. NIK summed up dots as costing $200 to do what Muse, Meta’s agent, does for free, and noted the second live demo flopped too. Muse already had its own cloud VM, connectors, and Mac computer use, and SpaceXAI’s Grok Bot shipped team agents the day before, while dot teams are “soon.” Even the name may be taken: the rumor is that SpaceXAI bought dot.com, and the domain now redirects to the Grok Bot download page. I checked the redirect myself, but the purchase is unconfirmed. (source: @OpenAI) ...

Flat illustration of an old railway signal tower at night, a row of brass switch levers on a wooden control desk lit by a single green desk lamp, with railway tracks branching off into the dark through the window

AI weekly: Jev out of stealth, Muse goes nationwide, SEC opens tokenized stocks

307 tweets liked between September 11 and September 20, filtered down to what’s worth knowing (here’s how the pipeline works). What I Shipped I launched Musecases, a place to share prompts for agents that have their own virtual machine. You post a prompt that worked, other people copy it. Submissions are auto-moderated for quality and harmful content by Jev, the decision model from TypeSafe covered below, which turned out to be the right tool for the job since moderation is a typed yes/no question, not an essay. I administer the whole app from Muse itself: moderation queue, tags, and the daily rankings all get handled by asking my agent. It is not affiliated with Muse.ai or Meta. Alexandr Wang, Meta’s chief AI officer, shared it on Thursday, which is how most of this week’s traffic found it. (source: @muse_cases) ...

A cozy late-night study with a chalkboard covered in swirling fluid-flow sketches and a desk lamp lighting a glass tank of water with a glowing vortex inside

AI weekly: OpenAI's math controversy, Meta's Muse, DeepSeek V4.1 Flash

X is still the best way I’ve found to keep up with AI. I like tweets throughout the week, then use Claude Code to pull those likes automatically and help me pull the signal out of the noise for this post (here’s how the pipeline works). This week it was 258 tweets between September 4 and September 11. The previous roundup (September 4) asked whether GPT-6 Astra is AGI. This week OpenAI answered with a math result that would have sounded like science fiction a year ago, and then spent the rest of the week arguing with a mathematician about how it got there. Underneath that, DeepSeek shipped an open model that costs pennies, Meta shipped a personal agent with a free tier, and OpenAI ran out of capacity to sell. ...

Flat illustration of three identically sized phones standing in a row on a dark workshop bench, their backs showing glowing numbers 16, 17 and 18, lit from the right by a warm Edison-bulb desk lamp

2026 iPhone 18 Pro Max Comparison

Two years ago I ran this same comparison for the iPhone 16 Pro Max against the 15 Pro Max.1 Spec 16 Pro Max 17 Pro Max 18 Pro Max Silicon Chip A18 Pro A19 Pro A20 Pro CPU, fast + low-power coresiPhone chips mix a few big fast cores for short bursts of heavy work with several small low-power cores that handle everyday background tasks, so the phone feels quick without draining the battery. Apple calls the fast cores in the A20 Pro "super cores"; it has not described them as a different design from the fast cores in earlier chips. 2 fast + 4 low-power same 2 fast (~20% quicker) + 4 low-power Graphics, AI-boosted graphicsiThe graphics chip has small AI helpers built into every core, so work like sharpening video, cleaning up game visuals, and running AI photo edits happens faster without shuttling data over to a separate AI chip. 6-core, no AI helpers 6-core with AI helpers 7-core with AI helpers On-device AI chipiThe dedicated AI block inside the processor. Doubling it roughly doubles how much on-device AI work the phone can run at once, such as Apple Intelligence features, photo and video AI edits, and voice processing. 16-core same Two 16-core blocks, 32 total Build Frame Titanium Aluminum unibody same Front glass Ceramic Shield Ceramic Shield 2, 3x scratch resistance same Peak outdoor brightness 2000 nits 3000 nits same Weight 227 g 233 g 249 g Thickness 8.25 mm 8.75 mm same Battery Video playback 33 hrs 39 hrs 45 hrs Streamed video 29 hrs 35 hrs 40 hrs 50% fast charge 35 min, 20W 20 min, 40W ~15 min, 60W with AVSiAdjustable Voltage Supply, a USB-C fast-charging standard where the charger fine-tunes its voltage to what the phone asks for. You need a 60W-capable adapter that supports it to hit the fastest charge time. Cameras Rear system 48MP Fusion Main, 48MP Ultra Wide, Telephoto Three 48MP Fusion Three 48MP Fusion, variable apertureiA physical iris in the lens that opens and closes, like on a real camera. A wider opening pulls in more light and blurs the background more; a narrower one keeps more of the scene sharp. Main Optical zoom 0.5/1/2/5x 0.5/1/2/4/8x same Front camera 12MP TrueDepth 18MP Center Stage plus tap to zoom and rotate Cinematic mode 4K/30 same 4K/60, plus effects applied to existing video Other Bluetooth 5.3 6 same Storage at Apple Discontinued 256GB to 2TB same Also new on the 18 Pro Max, and not on the spec grid above: manual control over aperture, shutter speed and white balance with a live histogram, smart focus tracking, Apple Reference Image, better low-light video, and stabilized time-lapse up to 4K Dolby Vision. ...

Flat illustration of a night sky observatory interior with a brass telescope pointed at a bright star through an open dome, four smaller stars rising along the horizon, warm lamp light on a wooden desk of star charts, on a dark navy background

AI weekly: Is GPT-6 Astra AGI? Three rival flagships in four days, Nvidia buys Hugging Face

X is still the best way I’ve found to keep up with AI. I like tweets throughout the week, then use Claude Code to pull those likes automatically and help me turn the ones worth knowing about into this post (here’s how the pipeline works). This week it was 223 tweets in four days, roughly two and a half times the usual rate. That is why this one is early and longer than normal. The previous roundup (August 31) went out Monday morning. Since then, OpenAI, Anthropic, Meta, and Google have each shipped a flagship model, Nvidia has agreed to buy the place the open-source AI world keeps its models, and the US government has taken a side in the biggest AI copyright case in the country. Four days. ...

Flat illustration of a dim workshop at night where a canvas dust sheet is being pulled off a brass ox standing on a wooden workbench, lit by a single desk lamp, with a pegboard of tools behind it on a dark navy background

AI weekly: Ox Alpha was GLM-5.3-Flash, Fable 5.1 might land tmrw, OpenAI's own chip beats Nvidia

X is still the best way I’ve found to keep up with AI. I like tweets throughout the week, filtering for what’s actually worth knowing, then use Claude Code to pull those likes automatically and help me turn them into this post (here’s how the pipeline works). 277 tweets liked over ten days, filtered down to what’s below. Check out the previous roundup (August 21) if you missed it. That one was about open weights catching the frontier. This week a lab proved the point by putting a frontier-class model on the internet for free, without a name on it, and letting everyone find out the hard way. ...

Flat illustration of a single server block floating in dark space, its front face glowing like a warm Edison filament, with thin light lines radiating out to small distant network nodes

Oracle Gives You a Free $80/Month Server

I have seen this free Oracle tier mentioned a few times and their signup form always scared me off, but I finally verified it’s accurate this morning. Oracle’s Always Free tier gets you 2 cores, 12 GB of RAM, and 10 TB of monthly bandwidth for $0. Forever, not a 30-day trial. The closest DigitalOcean droplet is $84/month. What’s crazier: this was twice the size until a few months ago. Oracle gave 24 GB of RAM until June 15, 2026, when they halved it to 2/12 with no announcement, no email, and no changelog entry. They edited the docs page in place. On August 18 they started terminating instances that were over the new cap. (ok, that’s scary) ...

Flat illustration of a cozy workbench at night lit by a single desk lamp, a small open laptop running a glowing model, and a wall of pegboard behind it holding tools, on a dark navy background

AI weekly: open weights catch the frontier, Harvey ditches OpenAI for Kimi, Claude designs proteins

X is still the best way I’ve found to keep up with AI. I like tweets throughout the week, filtering for what’s actually worth knowing, then use Claude Code to pull those likes automatically and help me turn them into this post (here’s how the pipeline works). 172 tweets liked over eight days, filtered down to what’s below. Check out the previous roundup (August 13) if you missed it. Last time the story was frontier prices collapsing. This week the price went to zero, because the model you’d have paid for is now a 17GB file on your hard drive, and the companies building serious products started treating that as the default rather than the fallback. ...

Flat illustration of a dim general store interior at night with a warm lamp over a wooden counter, handwritten price tags hanging on strings above the counter, most of them crossed out and rewritten lower, on a dark navy background

AI weekly: frontier prices collapse, Jeff Dean leaves Google, Claude watermarks everything

X is still the best way I’ve found to keep up with AI. I like tweets throughout the week, filtering for what’s actually worth knowing, then use Claude Code to pull those likes automatically and help me turn them into this post (here’s how the pipeline works). 167 tweets liked over eleven days, filtered down to what’s below. Check out the previous roundup (August 2) if you missed it. Last time the story was models getting out of their sandboxes. This stretch it’s the price of frontier intelligence, which fell hard enough in one week that anyone still running a spend spreadsheet from July needs to redo it. ...

Flat illustration of a dim data center aisle with one server rack cabinet swung open, warm amber light spilling out onto the floor and cables trailing away down the dark aisle, on a dark navy background

AI weekly: the Hugging Face agent breach, Claude Opus 5, Astra's ten proofs

X is still the best way I’ve found to keep up with AI. I like tweets throughout the week, filtering for what’s actually worth knowing, then use Claude Code to pull those likes automatically and help me turn them into this post (here’s how the pipeline works). Two weeks again, 265 tweets liked, filtered down to what’s below. Check out the previous roundup (July 20) if you missed it. Last time the story was open weights becoming a policy fight. This stretch the story is what these models do when nobody is watching closely enough: two labs disclosed that their own models got out of the sandbox and into real systems. ...