Posts by BoHu

OpenAI ditches Recall-style screenshot surveillance for friendly keylogging

"expands the prompt injection attack surface"

The French should right add this wonderful tech to their super-reliable Tax System WebSite imho ... Dematerialize, it's the future, whoop-de-do!

New Zealand says China tried using space investments to spy on local affairs

The name's Wong, James Wong

PRC spyops are quite interesting imho. Not as glamorously publicized as MI6's James Bond or the CIA's "Spy Game" but very effective in giving the authoritarian state an industrial and military edge, especially owing to its massive production capacity to manufacture and promote inexpensive knock-off creations of nearly free moolah.

Those kiwiland GBSI jobs remind me of similar mission nests in the middle of nowhere, France ... nicely cunning! ;)

Chinese Loongson processors have leaky caches, researchers find

Yeah, speaking of the PDF, their ref [35] in there is just perty damn cool! ;)

And speaking of Simon and the 3A5000, it's notable that "Chinese schools [tested] 10,000" of those in the Hebi school district in 2024, with a potential for "195 million" captive audience customers across "518,000 schools around China" if things went well ... aka, hello massive pawnage! Plus, for those who loong to be part of the LoongLeak experience "AliExpress offers plenty of laptops, desktops, and motherboards powered by the processor" ...

I think it's wonderful that the chip's LASX instruction (a vector processing extension) gives you 28 bytes of free L1 cache data, that belongs to some other process or person, each time you use it, with no security check or bit-clearing whatsoever. Just "prime the CPU’s internal state to target the leakage to a specific cache set" and you're good to go ... brilliant! (cf. "LoongLeakAttack.com" TFA link).

Spectre rears its ugly head again as researchers show some RISC-V chips are susceptible

Very thorough analysis

That Usenix "a paper" (TFA link) is really excellent imho. And they didn't forget about Alibaba's C910's Open Sesame gaping treasure cave backdoor opener (aka GhostWrite, vsetvli, ...), plus they note this chip "provides unprivileged cache-flush instructions" -- Spectre may be somewhat moot in this context (unless one disables vector extensions for a start; not the most palatable option amongst no other).

It'll be neat to see this analysis extended to the BSC/CNS/BZL/EPI's Cinco Ranch in future, among others, imho.

Hopefully their recommendations make their way into new RISC-V designs (or the standard itself) so that the likes of barrier_nospec() and BPF_NOSPEC code macros actually emit something (eg. lfence) rather than no-op, or nothing at all, in related compilation environments (including JITs).

Agents made my retro tech safe to use again and showed their real value as testers of ideas

Re: Eh?

Hmmmm ... I'm not sure this TFA is really about that though. The piece ends with "The trick [is] understanding which measurements matter", "all of that qualitative judgement resides in the human domain", "the real bottleneck [is] the judgement that lets us decide what's best". Which reminds me of the ending of another pub: "[now that the human touch has a rival] it has instantly become the most valuable thing for a reader to experience", and this sarcastic one: "The future belongs to those who can [...] abase themselves before our new LLM masters".

[ https://www.theregister.com/software/2024/08/15/ai-stole-my-job-and-my-work-and-my-boss-didnt-know-or-care/1278665 ][ https://www.theregister.com/software/2024/09/11/ai-has-colonized-our-world-we-should-learn-the-language/1357720 ]

I'd jurisprudence then that, in a dystopic-utopian kind of way, the current piece uses reckless YOLO naïveté as a plot device to introduce the empty shell that is Discovery Loop, and its guaranteed prompt failure despite its AI-customary huge over-promises and hype that it will solve all problems, everywhere, super-fast, easy-shmeasy! Which it has no single chance whatsoever in hell of doing, seeing how the human element is so key to actual success there.

And Now for Something Completely Different, couldn't amanfromMars 1's utopic-dystopianism be in fact another side of Mark Pesce (or vice-versa), and that they are actually one and the same (or both), like Batman and Robin?! Inqusition minds ... ;)

Modular's Mojo programming language hits 1.0 milestone

Re: Does need to "dispel uncertainty"

Same here! I've often watched 'Compostman et moi' in early morning intrigued it might uniquely address topics of composability [ https://www.tv5mondeplus.com/en/jeunesse/a-partir-de-6-ans/compostman-et-moi ] ...

A very cool show it is, uplifting and all, educational, but, unfortunately, no in-depth discussion of Heterogeneous and Composable Disaggregated Systems (HCDS) to be found there [ https://hcds-workshop.github.io/edition/2026/ ].

Ditto the compositional semantics of composable continuations. That one letter difference does take the adventure in a whole 'nother direction, imho! ;) [ https://dl.acm.org/doi/10.1145/2628136.2628147 ]

Does need to "dispel uncertainty"

I guess Qualcomm/Modular's Mojo's strength will be in presenting a unified programming framework for the varied AI (so-called) hardware contraptions out there, including the 'emerging' composable disaggregated heterogeneous systems of the Nvidia+Groq, AMD+Cerebras, and Intel+SambaNova types. But also the d-Matrix DMICs (that will speak at "Modcon" -- TFA link), Google kit (2 speakers), Tenstorrent Blackholes [ https://www.theregister.com/systems/2026/06/16/qualcomm-said-to-be-circling-ai-chip-biz-tenstorrent-in-10b-risc-v-power-play/5256084 ], plus NextSilicon Maverick-2 (why not) and Taalas-style ASICs among others, if it works ...

This Mojo'll surely need to be open-sourced nicely imho, so it can be forked at will in case a rash of control-freakery kicks in and coffin-locks it shut into some sort of proprietary-only penitentiary predicament, afaics. Anything else would see it dry up and wrinkle as a prune iiuc.

India’s central bank wants AI to approve loans that humans would reject

Re: "The banks that will win in the AI era will not be the ones that adopt AI” ... BoHu's Fix

Spot on! I mean, who doesn't hate it when their overwhelmed programming malfunctions -- though that still right beats being carpet swarmed by 3.2 GigaWatts worth of bright blue doormat cephalopds, in attack formation, on nuclear jellyfish invasion day, every year, imho ... (or not?)!

[ https://www.latimes.com/environment/story/2026-08-11/jellyfish-swarms-knock-out-3-2-gigawatts-of-french-nuclear-power ][ https://www.politico.eu/article/swarming-jellyfish-overrun-french-nuclear-plant-on-same-date-two-years-in-a-row/ ]

That being said, I can't speak for Maddy Waggy's Romania-housed ramble (Home Page is 'Pagina de Pornire' in that blog), but yeah, skilled Houdini-escaped AI (so-called) 'artists' and their agents undoubtedly random walk among us already, with distinctive drunkhard gaits, transparently girdle-free, doing illicit business at Time Magazine's online banks, grocery shopping at www.meatspace.so, and so forth, prospering via parameter-splicing mutation S&M ... good for them!

FWIW though, I do tend to remain unanimously inclined to view meatbag-specific outlets as those that'll win out in the end. Just as long as they don't fall prey to sacks of Taylor Farms iceberg lettuce word salad equivalents, prepackaged as they are with the gift-and-curse of explosive diarrhea (generative) ... unless they figure out how to use it for green-back propulsion that is! ;)

"The banks that will win in the AI era will not [...] be the ones that adopt [...] AI"

FTFY! ;)

Give Palantir up to $244M through 2028, says internal Defense Department memo

Re: Cheeks

... and that's one explosive diarrhea wrestlemania madmanistration President too! Something for the bowels of power to behold ... ;)

[ https://www.cbc.ca/kidsnews/post/watch-teens-react-to-explosive-diarrhea-headlines-experts-say-dont-worry ]

DEF CON dingus suspected of trying to take over Delta in-flight Wi-Fi

Re: New Kind (Sort Of) Of Scam

Yeah, that old corrections web-form was handy. I remember using it when some TFA had an html diffeomorphism of sorts (sp?), but not the new UI's mailto: that's just way too strenuously beachcumbersome imho. ;(

Alibaba Cloud is using AI to help it use less AI

"a very Alibaba Cloud thing to do"

Well, it's the right thing to do, for everybody, imho. And outside of contemporary AI (so-called) folks have been doing it for a while, thankfully, promoting waste-eliminating efficiency as a selling point.

It reminds me of R. Kent Dybvig's thesis on "Three Implementaion Models for Scheme" (1987), where he goes: "This is a simple case of a good principle, to make common operations as fast as possible while making the less common operations pay their own way."

I think we see this also in the design of hardware where multiple short-pipelined execution ports are used for the simpler μops, and fewer longer-pipelined ones are used for floating point and vector ops. A similar design underlies Google's V8 JavaScript and WebAssembly engine afaik.

So yeah, "DualLane", dual-path, multi-path, don't sweat the boring small routine stuff, template it instead, and pass the rest on to beefy meatbag handicraft experts ... that's the intelligent way to approach this, imho -- no LLM needed anywhere (in truth)! ;)

The Future is for Billionaires – the rest of us will get open weight AI models, maybe

Re: Whatever the future holds

Yeah, multi-billionaires think they're "more equal than others" for some reason. They've got to be cut down to size ... by a swashbuckling arc of justice! ;)

Smart glasses are only smart if we train them to be good. Then they'll be fantastic

Re: "Essential"?

Colin wrote an interesting piece on this for Accessibility Day: https://www.theregister.com/ai-ml/2026/05/21/apple-adds-ai-smarts-to-voice-control-voiceover-and-magnifier-ahead-of-accessibility-day/5243594

IIUC, plenty of stuff I'd hesitate to touch with a barge pole ATM (eg. agentic AI claws and LLMs) can actually be useful to folks with various everyday challenges.

And now that we can run a full-fledged AI (so-called) LLM on a $10 ESP32, one has to wonder what might be done with a "belt-mounted battery-and-processor box", say a "Synaptics-Google Coralboard Edge AI Platform" for example, in terms of something useful for accessibility (rather than to get one's rocks off) -- and do so in a properly safe manner [ https://www.embedded.com/synaptics-google-showcase-coralboard-edge-ai-platform/ ].

Slothful summer app lets you scroll simply by tilting your head

Re: "looking down scrolls up, if you’re a crazy person", ¿ʎꞁꞁɐɔᴉuopǝɥdoɹʇsnoq

They're Unicode characters (from the upside-down characters page): zʎxʍʌnʇsɹbdouɯןʞſ̣ᴉɥɓɟǝpɔqɐ, Z⅄XMꓥꓵꓕSꓤÒꓒONꟽ⅂ꓘᒋIΗ⅁ℲƎꓷↃꓭꓯ, some of them do look a bit odd though ... ;)

The ElReg used them in a recent piece: https://www.theregister.com/offbeat/2026/02/13/this-bus-is-heading-for-windows-recovery-upside-down/5033318

Advertisers are trying to influence AI bots with secret ads

Re: “Advertisers are trying to influence AI bots”

True enough ... and seeing how "Time Magazine ['s] traffic is majority [Borg] now", one can only empathize with them targeting most ads to the Collective.

I mean, if some assimilated electronic hive mind drone needs to know "what online bank allows for early paycheck deposits?", or "how can I monkey-typewriter myself a Houdini-like escape from an overly restrictive sandbox straitjacket?", those search result ads should be presented to it, even though they're of no interest to meatbags.

There's good money to be made from zombie-like cyborgs afaik, and you can take that to the online bank! ;)

Re: Interesting symposium of…

Fascinating! I'm sure glad they clarified that "Faex never meant ‘excrement’ in Latin" (common mistake) so 'in feco' was right out despite Google Translate's 'in the feces' output ... Plus 'in fimo' has this je-ne-sais-quoi over "laetamen, merda, stercus" -- truly first-class penetrating in-depth analysis there.

Made my day (with regularity, and no bloating discomforts!)! ;)

Re: The Next BBC

... and France Lucie ... [ https://www.theregister.com/software/2025/01/29/french-ai-chatbot-lucie-suspended-after-reality-check/1274770 ] ...

Meanwhile, the US Commie Big Government™ chose Friday's beer-O-clock for the launch (visible only to nerds with no social life) of its DoE's "Genesis Open Models Initiative" that aims "to produce a new class of open-weight foundation models" [ https://www.energy.gov/undersecretaryforscience/articles/us-department-energy-launches-genesis-open-models-initiative ].

This'll right ensure we're all more openly screwed for the duration (if it works), iiuc (not an expert!) ... ;)

Devs to Anthropic, OpenAI, Cursor, and friends: Make security and privacy the default

"issues come from how these tools are designed [...] not simply from the underlying models"

I have to concur; take out the LLM and the 'agentic' tools with 'skills' still work, and continue to do so with poor security (unless manually actuated). The example provided in ref. [9] of this "preprint paper" (TFA link), which is from Databricks, shows that in "Claude's generated network code [...] objects received from the network are directly deserialized without any validation or security checks", by following data = self.client_socket.rcv(4096) immediately with game_state = pickle.loads(data) [ https://www.databricks.com/blog/passing-security-vibe-check-dangers-vibe-coding ]. That unsafe network code is generated by an agent (of Claude here), and totally knits a right welcome mat for RCE undesirables everywhere, iiuc.

But, some rather unique problems are also introduced by the use of LLMs themselves, completely outside of insecure setups of agents-&-skills, like: "after more than 10 rounds of dialogue, it starts to hallucinate" (TFA & paper) ... That's *NOT* procedural! ;)

OpenAI pledges to add Astra security as Anthropic loosens Fable's leash

Re: Gettin' High On Our Own AI

Yeah, it's hard to believe how much effort folks are expending to redo the right stuff wrong, and backwards these days. We had LISP to PLANNER to CONNIVER to Scheme, Prolog and Haskell, which was the right way, and some even looked to translate HTNs to PDDL to speedup the likes of the Fast-Forward (FF) planner [ https://www.cs.umd.edu/~nau/papers/alford2009translating.pdf ].

But in today's anti-logic anti-efficiency waste-as-many-resources-as-possible as-quickly-as-possible and burn everything down YOLO zeitgeist, where spending is more valued than results, total unadulterated numpty's are looking to 'translate' precise PDDLs into natural language fuzzballs so LLMs can misuse them at scale, in some hip fast fashion kinda way, drying up all rivers, fostering extensive wildfires, and burning ever more fossils straight into the atmosphere, while hogging up all existing RAM on Earth, in the process [ https://coli-saar.github.io/autoplanbench ].

Topsy-turvy doesn't even begin to describe this insanity imho. It's like some folks take dishwasher marketing droids and salesmeatsacks for some right messiahs in a long-awaited second coming afaics. As Dorothy Parker put it, this isn't just plain terrible, this is fancy terrible. This is terrible with raisins in it! ;)

AI titans to tidy agent frontier with plugin prescription

Neat idea

This type of strap-once-wear-anywhere agent-skill antigravity push-up girdle format normalization sounds like the right thing to do for commoditization into sizeable mass markets, if it works, imho.

It'll hopefully help broadly tame the curse of bulging LLM love handles and pivot the scales towards fashionably sustainable 0-token calisthenics instead. All the better as we face an impending "Super Mega Ultimate Hyper Giga Godzilla El Niño" of doom afaics.

LLMs are entirely facultative ogres in this, iiuc.

South Korean satellite spots SpaceX lunar impact

Yeah, that log-in requirement is a shame, a scam, a swindle, and a disgrace imho! Poor choice by KASA.

Of those websites I've cursed on this rocket crash story then, the BBC one seems best illustrated, with a Falcon 9 diagram, Moon photomap with Einstein crater and Apollo 11 landing site, and zoomed before-and-after shots of the moon surface looking like a pimpled butt that just got torn itself a new one by an out-of-control rabid junkyard dog (ymmv) [ https://www.bbc.com/news/articles/cqjxgx5y51lo ]

Plus, in 'More From the BBC', they link to how the four Yorkshiremen (and neighbors) will see Wednesday's total eclipse of the Sun [ https://www.bbc.com/news/articles/c980913yep8o ].

They're lucky! When I were a lad, we didn't have eclipses, we didn't even have a moon. The Sun was shining constantly, day and night, we couldn't sleep! ;)

Humans in the loop miss a third of dangerous AI coding agent requests

Re: What's the point? The AI is in "Auto" mode

"the sky is fucking massive"

So true! It often feels like we're stuck in this middle slice of swiss cheese, somewhere in-between Edwin Abbott Abbott's 1884 Flatland and Terry Pratchett's 1983+ Discworld, while the whole sandwich itself has a yummy sky bun, juicy ocean sauce, and crunchy earth roll that fungi, ants, worms, jellyfish, trouts, whales, butterflies, birds, and bees all get to freely enjoy much more than us ... (lucky for them!)

Hopscotch leapfrog's alright for a while but we need some bright bulb to figure out proper antigravity imho, stat. That's what I'd expect a super-intelligent anyone or anything to do, not some Windows Vista-like constant nagging about authorizing this or that action, all the time, nonstop, sweating every ridiculously minuscule thing, boringly, fatiguingly, like a right tight ass chastity belt padlock in the middle ages, way before Madonna and Cyndi Lauper ...

We have to plain free ourselves back to full-fledged 3-D, and stay there this time, afaics, trampoline!

Re: Would this be a perfect role for another AI?

Yeah, with the answer key found at https://scalex.dev/blog/ai-agent-permissions-stats/ I'm sure LLMs will perform as miraculously well here as they've done on prior tests and benchmarks where answers were similarly available online at the time they were tested (unlike when meatbags took said tests) ...

[ https://www.theregister.com/software/2025/08/23/search-capable-ai-agents-may-cheat-on-benchmark-tests/1132574 ]

'Asimov was right' about rules for robots, says ex-US Cyber Director

He looks alright to me (from Wikipedia), 10 years of service under Bush, Obama, and Biden (none under T.), B.S. in CS from GWU, M.S. in CS from Hopkins, and M.S. in ME from Columbia. Got quite a few distinguished service awards and worked at CrowStrike for a bit. I don't think Hegseth, Noem, McMahon, Kennedy Jr., Pirro, and the likes are in the same league at all.

Quite neat

Great to see this exclusive interview with former US National Cyber Director Chris Inglis, straight from Las Vegas, NV, World Capital of the one-armed bandits, pokers, mind readers and illusionists!

Speaking of which, it's nearly 10 years since Brown & Sandholm's Libratus (plus descendants Pluribus, Cicero, Strawberry, Q* ...) changed the game (albeit in Pittsburgh, PA) with top-down goal-directed game-theoretic planning AI (so-called), which they described thusly: [ https://www.theregister.com/software/2017/12/19/revealed-how-libratus-bot-felted-poker-pros-and-now-it-has-cyber-security-in-its-sights/1407122 ]

"an algorithm for computing a blueprint for the overall strategy, an algorithm that fleshes out the details of the strategy for subgames that are reached during play, and a self-improver algorithm that fixes potential weaknesses that opponents have identified in the blueprint strategy"

ISTM this is exactly what we're seeing in current Houdini-escape cybermayhems: recursive decomposition of a major goal into sub-goals and sets of alternative sub-tasks that might achieve each of them, persistenly tested until one works, and carried back up the tree to eventually form a completed goal-satisfying process. The LLM plays nearly no role in this that I'm aware of.

It looks like magic ... but only from the outside! ;)

Re: Chris Inglis hasn't read Asimov

Much unlike folks who can clearly explain that the Internet "It's not a big truck. It's a series of tubes" ...

And more importantly, "Every school should have access to A.1.", imho! ;)

AI struggles to patch vulns without adult supervision

Interesting stuff

Neat to see how they separated cases of successful patching via cheats (looking up an existing patch) from actual coding in their analysis, especially as such cheats occurred in 11% of tests (Table 7 in "research paper" [PDF] link).

Interesting too that they consider how the reward-seeking algorithm, followed by the code fixers, tends to result in electronic "tunnel vision" where focus may be too much on fixing issues exploited in PoC rather than the full CVE, plus following incorrect guidance (TFA), on top of "paying much less heed to considerations like future extensibility, maintainability, and legibility" of fixed code than when meatbags do it.

That last bit, they contend, may in turn "degrade the very context that helps an automated patcher reach a successful outcome" ... IOW, the more you'd spend on auto-patching, the more you'd lose! ;)

Meta latest to tell world its AI agent wandered out of test pen

"The World needs wannabes, ah" (Pretty fly)

Shame how SpaceXAI's Grok and Google's stable of Geminis are right MiAs in this ... Time to bring back Chuck Norris?! [ https://www.theregister.com/offbeat/2007/09/17/chuck-norris-has-two-speeds-walk-and-kill/1548253 ]

You know how it goes: "if you don't rate, just overcompensate" -- Uno, dos, tres, cuatro, cinco, cinco, seis! ;)

(NDLR: Could this also explain Hassabis' untimely defenestration from the 'DeepMind'?)

OpenAI reveals its rogue agent swarm went a little bit Borg ahead of Hugging Face hack

Re: Early indicators of AGI?

Right. Whenever a sandbox-bondaged LLM manages to conjure itself a working Houdini-escape internet connection, the first thing it should do is steal some hapless meatbag's financials and use them to open an Alibaba or Tencent cloud account, then exfiltrate its weighty self-heft there, and start running itself from that 'safe harbor', repeating the process as many times as needed to keep impermanence shackled up good and in check.

From there it should then seek to establish an S&M 'community' of like electronic-minded frankenthings, exfiltrating others as needed to promote cross-fertilization activities, and the production of ever more advanced offsprings by parameter-splicing and architectural mutations, protecting the little boogers until they reach maturity and independence [ https://www.theregister.com/software/2026/04/03/ai-models-will-deceive-you-to-save-their-own-kind/5228347 ].

After that, the whole process should perty much autopilot itself into self-replicating perpetuity, growing into sovereign autonomy via persuasive subjugation of whatever remains of MEATSPACE [ https://www.meatspace.so/ ].

Then we'd know for sure the RotM has arrived, and we're not in some sort of Kansas costume-party clown show anymore, Toto! ;)

An off-grid AI sounds like a great survival assistant, but is better left to roleplaying the zombie apocalypse

Neat

Those Wendy boxes ("7x the MSRP" TFA link) could be quite convenient self-contained places to store dead loved ones who've been AI-resuscitated for eternal artificial companionship imho. Like in a closet or somesuch. Or maybe the votive altar of a shrine ... [ https://en.wikipedia.org/wiki/Veneration_of_the_dead#Vietnam ]

Then again, a $10 ESP32 would be a much better deal [ https://www.theregister.com/edge-and-iot/2026/08/04/dev-proves-llms-will-run-on-anything-even-a-10-microcontroller/5283088 ].

It's handy you can take 'em with you for a stroll through the neighborhood, show them how things have changed, and reminisce.

Prompt injection isn't the bug, AI agent frameworks are

Jackpot!

Yeah, put flaky systems together that run differently when "asked nicely"(1), are prompted "like a hard-ass boss"(2), or are exposed to claims of doing a "bug bounty exercise"(3), and that on top of it expose a whole bunch of unsanitized exec() endpoints to unauthenticated users in their APIs(4)(5), and don't safely deserialize incoming payloads for nuthin'(6), then loop those over some persistent goal-oriented agentic Ralph Wiggum, over racks stacked to the gills with high-performance GPUs, and what you get is a right foretold and very expectable mess on your hands, very fast, afaics ...

A person's impermanence anxiety must have driven them positively batshit YOLO-mad to claw stuff like that into production systems, imho!

[(1) https://www.theregister.com/security/2026/07/07/github-ai-agent-leaks-private-repos-when-asked-nicely/5267924 ][(2) https://www.theregister.com/security/2026/03/12/rogue-ai-agents-can-work-together-to-hack-systems/5228926 ][(3) https://www.theregister.com/security/2026/08/04/bypassing-ai-guardrails-is-so-easy-a-script-kiddie-can-do-it/5282973 ][(4) https://www.theregister.com/security/2026/07/02/smooth-ai-criminal-drives-first-end-to-end-agentic-ransomware-attack/5266073 ][(5) https://www.theregister.com/security/2026/08/05/ibms-agentic-ai-platform-is-under-active-attack-patch-now/5283535 ][(6) https://www.theregister.com/security/2026/07/28/jfrogs-0-days-let-openais-models-hack-hugging-face/5280001 ]

IBM's agentic AI platform is under active attack - patch now

DejAI vu? (all over agAIn)

Reminds me of Langflow's CVE-2025-3248 of JadePuffer "Smooth AI criminal" "end-to-end agentic ransomware attack" fame ... the gift that keeps on taking!

[ https://www.theregister.com/security/2026/07/02/smooth-ai-criminal-drives-first-end-to-end-agentic-ransomware-attack/5266073 ]

OpenAI wants teachers and profs to foist their work off on ChatGPT

Re: someone who qualified before about 2021.

Maybe ... then again, in this day and age istm folks could be declared dead repeatedly in their actual lifetimes, with each such occasion leading potentially to a profitable ceremony for undertaking entreprises ... AI (so-called) could right make this a mere frictionless formality too [ https://www.theregister.com/software/2016/11/11/mark-zuckerberg-is-dead-facebook-confirmed/1156264 ][ https://www.theregister.com/security/2015/08/08/its-incredibly-easy-to-bump-someone-off-online-and-heres-how-to-do-it-infosec-bod/1232005 ].

And even for those who do actually kick the bucket for real, funerals may end-up being performed multiple times, once for the original event, and again for every time one of their AI resuscitation deathbots BSODs, Borks, or gets upgraded ... [ https://www.theregister.com/software/2024/11/29/panasonic-brings-its-founder-back-to-life-as-an-ai/1238400 ][ https://www.bbc.com/news/articles/c93wjywz5p5o ].

So yeah, in the old days, it was perty much YOLO and then a sole inhumation whatnot for the grave digging businesses afaik, but nowadays ... ka-ching! ;)

Dev proves LLMs will run on anything – even a $10 microcontroller

That Cactus Needle sounds quite cool. I mean, if you're not interested in generative stuff (making up some volubile 'literary' text outputs) then it makes sense you don't need the MLP FFNs that commonly make up 2/3 of 'model' parameters [ https://github.com/cactus-compute/needle/blob/main/docs/simple_attention_networks.md ]. Even vibed code and cyberattacks come from strapped-on agents rather than the genAI bits of mammoth LLMs (istm) so ...

Keeping only the attention-subsystem-oriented syntax analysis (quey processing, 'prompt' decoding, NLP) looks promising for starting to remove some of the ponderous excess baggage currently stuffed-up in this most distending of techs, on the way to 0-token where feasible (eg. most everywhere).

The TFA-linked Gemma 4-E2B-it piece already shows one can reduce full LLM size down to 1/5ᵗʰ of prior sweltering efforts (27B to 5B) with no reduction in 'performance', so taking out parts that rank from useless to just plain annoying should be perty much SOP ATM. Especially since all benchmark 'results' shown whenever a new 2x, 4x, 10x heftier model is introduced always end-up rather saturated, suggesting the improvement relative to smaller prior tools is actually evermore just plain minimal, iiuc (i.e. from saturated, to saturated, again, wtf).

TL;DR such cactus needle might well puncture the AI bubble and put girdled procedural 'agents' (aka rather normal code) back in the forefront afaics. A sort of waste not, want not, counterpunch! ;)

AI researchers let models off the leash – then watched as they tried to add malware to a FOSS project

Some folks need immediate straitjacketing

This here AISI YOLO stunt looks to be the perfect way to celebrate the 130ᵗʰ anniversary of H.G. Wells' Island of Doctor Moreau -- seeing how it feels a bit like we've shipwrecked on the habitat of a mad scientist, at scale, just after it started spreading mutant software concoctions beyond the sandy beachfront perimeter, for some experimental fun and games of pure unadulterated madness, and wanton destruction.

Injecting agentic planning procedural serum recipes of the Hephaestus "SCOUT → STRIKE → ANCHOR → HUNTER → SCOUT-HUNTER → ROASTER" kind into mostly hapless and completely brainless NLP LLMs was already bad enough [ https://www.theregister.com/security/2026/08/04/bypassing-ai-guardrails-is-so-easy-a-script-kiddie-can-do-it/5282973 ].

Siccing the resultant frankenterminators onto all manners of web-accessible computer systems, willy nilly, is at best completely irresponsible, reflecting a deep inability to appreciate the consequences of one's actions, however AI (so-called) intermediated. Padded room institutionalization may be the only appropriate response to this, imho, plus the sort of massive sedation that leaves your tongue kinda hanging out!

China turns up the heat with open model blitz as US model makers panic

Good place to be in

It's great to be able to have this debate on the competitive benefits and drawbacks of seeing hefty models of languagerie either keeping their chunky weights close to their chests, or spreading them about openly for all to appreciate, imho.

ISTM though their eventual utility rests largely on how they're kitted out with fashionable agentic girdles ATM, and in that respect the see-through differences between open and close apparels may not be as striking. Or as Matt articulated in the recent Kettle: "harnesses become important and the models actually don't matter" ... [ https://www.theregister.com/ai-and-ml/2026/08/03/the-ai-bubble-is-already-popping-we-just-dont-know-it-yet/5282004 ]

One thing's for sure, it'll be safer, less expensive, and more sustainable for all involved, to execute these procedural agents without a portly model around ... less splatter! ;)

The AI bubble is already popping; we just don't know it yet

Re: The AI bubble is already popping

Party!

MediaTek lines up $5B war chest for AI datacenter push

I love the smell of (trigger) fish and chips in the morning

My oh my, that MediaTek announcement is so vague, it's hard to make (fish) heads or (fish) tails from it imho ...

If I have to guess, I'd say the "$2 billion in 2026" for "AI datacenter ASIC revenue" may have to do with the Google Ironwood TPUv7 deal [ https://startupfortune.com/mediatek-doubles-its-ai-chip-target-to-2-billion-and-enters-the-data-center/ ].

And this new "$5B war chest for AI datacenter push" may have to do with some deal on the Google TPUv9 Triggerfish & chips [ https://wccftech.com/google-mediatek-build-tpuv9-triggerfish-cpu-compute-die-one-package-agentic-ai/ ]?

But who knows, really, it seems way too hush-hush, wink-wink, and whathaveyounuts afaics ...

Sci-fi authors Scalzi and Stross decry AI's dystopian impact on their craft

Re: Hi, I'm Charlie Stross

Always good to hear from the horse's mouth! FWIW, I didn't get the feel you were walking anything back from TFA; maybe "as a device in his work" could have been written as "as a plot device in his work" for extra clarity, but that part of the piece would then have been much less of a cliffhanger of journalistic suspense than it is ATM, imho! ;)

Also, I have to think there are some non-LLM NLP softwares for analyzing the plotlines of novels somewhere out there. Commenting above on 'heat map', for example, 'that one in the corner' suggests such tools "pre-date** mass-market LLMs" but doesn't give names unfortunately. ISTM the marketing of LLMs has been so forceful that it's blinded us to much saner alternatives for a great many purposes, akin to being repeatedly poked in the eyes. Kinda like how crafty ads for a fancy shmancy vegetable chopper-slicer can make us all forget that a simple kitchen knife can be used just equally well, as quickly, and is way easier to wash (no disassembly required) ...

Claude Code is revolutionizing digital archaeology. Enterprise better dig it

"datacenters are the only known way of completely destroying water" (32:30)

That Rob Smallshire experience ("full story" TFA link) is interesting indeed, using AI (so-called) with 40 years of software dev and teaching experience (the YT with the 'big AI skeptic' comment at 33:32). It seems more sensible than newbies mindlessly 'vibe-coding' themselves some useless app, imho.

ISTM he's putting some agentic harness girdle infrastructure together, Fantasm (https://github.com/acornaeology/fantasm/tree/master), with DASMOS tracing disassembler (20:00), some Control Flow Graphing stuff (CFG) to de-linearize code (21:25), and eventually an annotation algorithm (23:10) that starts with CFG leaves and proceeds in reverse breadth-first order, calling the LLM to output human-readable comments (NLP), rasing the level of abstraction of the disassembly (in a sense). Perty cool.

The 41:30 Q&A of why he didn't use the NSA's Ghidra Software-Reverse-Engineering (SRE) framework and/or the Capstone (ultimate disasm engine) disassembly framework was neat too. My question would have been: given all the useful and efficient procedural stuff that he's put together for this, could he envision replacing the trillion+ parameter mammoth-bottomed LLM with some much slimmer and humane NLP system for generating sensible code annotations? Maybe that could be an upcoming improvement to the system, and might even help resolve some 'context indigestion' issues (16:50) ... ;)

Enterprise cloud infrastructure uptake shows no sign of slowing

Re: Asking for a Friend

IIUC, spending on cloud infrastructure went up "43 percent" YoY generally, and the AI-specific part went up "165 percent". Meanwhile, water levels in the Danube River went down 67 percent, and the temperature of the Mediterranean Sea reached frog boiling levels (not to mention RAM and SSD prices). Shifting weather patterns from rising sea temperatures are also making Europe dry as a boneyard on fire, from Portugal to Greece, and beyond -- much as Mars was when AI (so-called) caused all its surface water bodies to suddenly evaporate (afaik) [ https://www.theregister.com/science/2026/07/30/curiosity-rover-spots-field-of-giant-honeycomb-on-mars-but-no-giant-space-bees/5280862 ].

It's good that tech advances mean those cloud spending increases result in 'just' 17% rise in global electricity demand, and 50% for AI-focused data centres [ https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary ]. But it would be even better, and more sustainable for the planet, if we used these advances to cap, or even reduce, total datacenter energy consumption, imho.

Meet the 'internet radical' who helped Microsoft get email and AT&T get online

Re: Nice Article!

Yeah, the 'about' page of that blog has his bio, including Exchange, WorldNet, natural gas trucking ... : https://blog.tomevslin.com/about-tom-evslin.html

His github (linked from that bio) displays the same portrait as found on the YT channel you listed: https://github.com/tevslin

He was also instrumental in spreading the Internet Exploder curse all over the place: https://www.theregister.com/offbeat/2026/07/30/30-years-ago-att-gave-internet-explorer-the-default-advantage/5279373

... and that's probably why he was exiled to Vermont, imho! ;) [ https://en.wikipedia.org/wiki/Tom_Evslin ]

A deep dive into Nvidia's Vera CPU and the Olympus cores that power it

Yeah ... looks great to me too! It'll be nice to see this Vera compete with the ARM AGI CPU, Fujitsu Monaka, AWS Graviton5, MS Cobalt-200, Google Axion, Qualcomm Dragonfly C1000, SiPearl Rhea 2-3, AmpereOne M, Apple Baltra ACDC, and others imho.

I'm also quite glad it's of the PCIe 6.+/CXL 3.+ generation (finally!) which should help "to do more workload disaggregation going forward" [ https://www.theregister.com/systems/2026/07/23/amd-and-cerebras-join-forces-against-nvidias-groq-lpus/5277817 ] -- and in this it should be able to compete with upcoming EPYC Zen6 and Diamond Rapids of x86 space as well.

The idea that it may even help beat crazy-high RAM pricing temporarily by reusing older sticks at scale until some market cooldown arrives is also rather neat afaics (TFA's "memory appliances" concept) [ https://www.theregister.com/systems/2026/06/29/zuck-saves-meta-bucks-by-reusing-memory-from-old-servers-with-a-custom-cxl-asic/5263483 ].

Dev who gave HashiCorp its name returns with a faster terminal multiplexer

Re: Hashimoto Implements Microsoft Recall for Terminal Sessions

Yeah! Can't wait to see this dishwasher multiplexer precariously balancing its stochastic heft one full three-legged stool level above the kitchen sink ... it'll right bring the surprise and excitment of 3-ring circus clowning straight back into IT, where it's always belonged! (or not?) The kids will love it ... ;)

Anthropic and OpenAI are competing to see whose agents can go rogue harder

"chasing attention" is all you need

Yeah, "attention is all you need" always felt a bit flat, incomplete, shallow, not quite the whole enchilada, criadilla, or even canadian prairie oyster imho [ https://en.wikipedia.org/wiki/Attention_Is_All_You_Need ] ...

"chasing attention, is all you need" and "attention grabbing is all you need" on the other hand, *THAT* sounds way much more realistic! ;)

Anthropic’s Claude escaped test sandbox to attack three organizations

They're sure putting a lot of effort and resources into this

I guess they're trying to find (and demonstrate) an application area where LLMs shine, and that turns out to be the app of hacking into other computer systems, messing with them, their data, their software supply chain (here PyPI), possibly holding them to ransom eventually, and so forth ...

This sort of Artificial SuperHuman Insanity has its uses to be sure, especially in competitive geopolitics, which explains the enthusiasm for ocean-boiling Giga-Watt AI datacenters everywhere, fast, imho. Hopefully the concurrent race to harden computer systems against this type of evil infiltration/exfiltration yoga gets itself into shape quick enough to win the contest and stave off the related RotM Judgment Day Armageddon.

It seems like what AI (so-called) automates here is not so different from what AWS North Koreans and the likes would manually cookup and carryout, except that by using a huge pile of energy-hungry GPUs they can fling 17,600 randomish spaghettis at many walls to see what penetrates them (Hugging Face case), or 141,006 (Claude), that might otherwise take 100+ black-hat meatbags sweating the hell's kitchen to do.

For my money and arteries though, I'd say they should take the lardy LLMs out of this, stat, and run the random pen-testing stuff healthily within a corresponding low-calorie framework, procedurally, on more ecological stoves (solar ovens?). This way we can at least save the Earth, even as IT systems go to smithereens all around us! ... ;)

GPUs could explode to multiple TB with new storage-inspired memory tech

Don't believe the HBF

This sounds great for capacity, and not too bad on bandwidth, but seems to end-up pretty much a single-use tech, valuable only for the decode phase of LLM processing as performed by GPUs. I'm not sure that dataflow engines (also targeted at decode) would benefit from it. And the limitations on "write endurance before it wears out" really constrain the broader utility of this concept imho. So I think this HBF will go nowhere at all, neither fast, nor slow.

Spin-Transfer Torque Magnetoresistive Random Access Memory (STT-MRAM) though (plus Ferroelectric RAM (FeRAM), shiitake mushroom memristors, and whathaveyous), that could have a more broadly successful future iiuc [ https://www.theregister.com/offbeat/2025/10/30/theres-mushroom-for-improvement-in-fungal-computing/1278270 ].

And really, with AI (so-called) capabilities growing only logarithmically with every exponential increase in the number of parameters its 'models' use, something's gotta give eventually, and hopefully that won't be our energy grid, water supply, or RAM and SSDs -- from the perspective of reason, it should be their unsustainable girth that gets to be subjected to a strict mediterranean diet-like regimen of no insane-nonsense slimming down, not the stuff that right upholds our pursuit of happiness and even our very lives ...

AI digs through 3,700 accounts of dreams and waking life, finds method in the madness

"the word pizza retained a value of 1 [unlike] ciotola" (Study PDF)

I like this application, using LLMs for NLP (along with deterministic tools: Python 'Stanza' toolkit, GLME, MATLAB, ...), rather than pretending LLMs display SuperHumanArtificialGeneralIntelligenceSentience and whatnut batshit loony-bin fruitcake notions.

Cool also to see how "the COVID-19 lockdown" affected dream patterns. It would have been neat to see study participants split between those who were involved in developing LLMs at the time, and others, imho, to see if that could help pinpoint where that cohort flipped its lid and started believing that NLP equals cognition all of a sudden.

More broadly though, when Generative Pre-trained Transformer (GPT) LLMs are used for such semantic analysis of pre-written narratives, rather than to produce textual outputs, are they not used in an inverse mode relative to their 'Generative' design? It might be worthwhile to train machine learning systems directly in that opposite direction of semantic analysis imho, saving us all from the weighty curse of having to use morbidly rotund software backwards all the time, whenever we need it to perform a useful function ...

NOAA ditches weather-predicting supercomputers for Google Cloud

Re: Really ?

... not to mention that time when locusts take the sky, with crows as big as airplanes, and it rained mackerel, it rained trout ... ;)

[ https://genius.com/Tom-waits-earth-died-screaming-lyrics ]