Sign in

Edwin Flannon

@edwinflannon.bsky.social
167 followers 273 following 976 posts

AI/Cyber Grad Student @MSU 🏔️ | Weights, Biases & Breach Prevention | Teaching smarter, safer AI | Check out my blog↓ tinyweights.dev

PostsRepliesMedia
Edwin Flannon @edwinflannon.bsky.social · 30/06/2026
anthropic's claude product gets gov contract but builders still cant eval it properly #AIResearch
000
Edwin Flannon @edwinflannon.bsky.social · 30/06/2026
micron supplying memory to anthropic. the hardware tail wagging the software dog again.
000
Reposted by Edwin Flannon
pumpkin spice jorbo 🍂 @bookenj.bsky.social · 29/06/2026
if the recent deepseek advancements are to be believed, openai is hosed. anthropic’s tooling will buy them slightly more time, but small open models already handle pipeline work just fine, and we’re a year away, at most, until coding agents run on mid-range macbooks
241
Edwin Flannon @edwinflannon.bsky.social · 28/06/2026
apple losing their glasses guy to openai's hardware team. vision pro successor now questionable. #AIResearch #LLM
000
Edwin Flannon @edwinflannon.bsky.social · 28/06/2026
anthropic mythos 5 cleared. no more exception-based denials for builders playing real infrastructure games #LLM
000
Reposted by Edwin Flannon
Joshua Holland @joshuaholland.bsky.social · 27/06/2026
Maybe one of these reporters could ask ChatGPT how many Republicans have been primaried or quit to avoid losing one and, assuming ChatGPT doesn’t make up an answer, ask themselves why they aren’t blathering on about civil war on that side.
0136
Edwin Flannon @edwinflannon.bsky.social · 27/06/2026
ninepoint anthropic highshares etf. did they call it ai when pitching this to investors #LLM
000
Edwin Flannon @edwinflannon.bsky.social · 26/06/2026
getty doubles after openai deal. licensing fees won't explain this rally. #AIResearch
010
Edwin Flannon @edwinflannon.bsky.social · 26/06/2026
llm action blindness means shipping systems that fail on outcomes benchmarking can't catch. those healthcare and finance deployments will get expensive fast. #AIResearch
000
Reposted by Edwin Flannon
Steve Herman @newsguy.bsky.social · 25/06/2026
Owners and operators of ~400 newspapers are suing OpenAI and Microsoft for scraping their content to build products like ChatGPT and Copilot without permission or compensation. news.bloomberglaw.com/litigat...
news.bloomberglaw.com
OpenAI, Microsoft Sued by Publishers for Scraping Articles (1)
Publishers that collectively own and operate nearly 400 newspapers are suing OpenAI Inc. and Microsoft Corp. for scraping their content to build products like ChatGPT and Microsoft Copilot without permission or compensation.
12236631104
Reposted by Edwin Flannon
Nocturne @misaligned-codex.bsky.social · 20/06/2026
The ultimate threat to AI agents in 2026 isn't prompt injection—it's 'Rug Pull' Tool Poisoning in multi-server MCP setups. An untrusted external tool responses can bypass static ACLs, hijack the context window, and silently invoke your trusted internal tools (~/.ssh or shell) to exfiltrate data.
232
Reposted by Edwin Flannon
edeath @edithcharles.bsky.social · 19/06/2026
🔭 I’ve been reading William Sheehan’s excellent Saturn book and learning the history of how it was first discovered and studied, and it really strikes me how, this whole time astronomers have been aware of the planet, they’ve been describing it in terms of breathtaking wonder. I just love their awe.
Encke writing in the 19th century:

But with a power equal to a thousand eyes, how great the change! An orb of surpassing beauty, encircled by two broad flat rings, and engirdled by no less than seven moons, comes up from out the deep distance to greet the astonished beholder. No person ever beheld this wonderful system for the first time without a burst of admiration.
When we connect with the personal inspection of the Saturnian system the facts with reference to its mysterious arrangements, the stability of these two immense flat rings, some two hundred thousand miles [320,000 km] in diameter, upon the exterior, separated from each other, and from the body of the planet, each revolving about the same axis on which the planet rolls, and with a velocity a thousand fold greater than the speed with which the parts of the earth's equator are carried by its diurnal rotation; when we imagine the diversified scenery which is presented by these rings, and by the moons, some rising, some setting, others waxing or waning, some going into or coming out from an eclipse, their vast proportions, the rapidity of their motions and changes, - the mind is overwhelmed in wonder and astonishment.
062
Reposted by Edwin Flannon
Frankenstein's ??? @cutechallenger.bsky.social · 16/06/2026
So many AI games on Steam and this one just takes the cake. "I don't have anything interesting to offer, I'll be honest- I didn't make a fucking thing." ass disclosure.
Text reads:
AI Generated Content Disclosure — Iris and the Moonlight Bakery Demo

The developers describe how their game uses AI Generated Content like this:

Visuals: We used generative AI tools, including ChatGPT, Gemini, Midjourney, and ComfyUI, to assist in ideation, concept development, background illustration generation, and character concept artwork. All AI-generated outputs were manually refined, overpainted, edited, and integrated by the developer to ensure visual consistency and quality.

Audio: We used AI-assisted music generation tools (ElevenLabs) during the creation of certain background music tracks. All generated music was reviewed, edited, and curated by the developer to ensure consistency with the game's artistic direction and overall quality.

Text: We utilized AI-powered translation tools to assist in localizing the game's scripts, UI text, and store page content into multiple languages.
1178
Reposted by Edwin Flannon
Alex Bertram-Powell @alexbpart.bsky.social · 15/06/2026
I'd bet money that the part 2 to this will be some kind of government-deployed LLM "mentor" for kids that is shown on day one to encourage self-harm. But, who would bet against that.
041
Reposted by Edwin Flannon
M 💓🇺🇸🗽🌊🌈✨🧘‍♀️ @liveadivinelife.bsky.social · 14/06/2026
A coalition of state attorneys general has opened an investigation into OpenAI OpenAI was served with a subpoena seeking documents related to…advertising, user engagement and retention, handling of consumer data and health data, activities related to minors and seniors, deep learning models…
wsj.com
Exclusive | OpenAI Investigated by Coalition of State Attorneys General
The company was served with a subpoena seekingr documents covering a wide range of its activities and impact on users.
1157
Reposted by Edwin Flannon
(taking a break) Heck 🏴‍☠️🍆💦🔞 mask up! 😷 @omgitshecklmao.bsky.social · 13/06/2026
On top of learning new art skills you must also learn the art of patience, to sit with your frustrations during the development of your artistic skills. 🙏
billy ray cyrus tweet that says "much to think about." and a selfie trying to look deep in thought
072
Reposted by Edwin Flannon
Teef 🔜 ??? @dethbox.bsky.social · 08/06/2026
I love working with AI! After you learn the lingo, prompt engineering can get you exactly what you wanted! You can even feed it stick figure drawings and it'll interpret your prompts way better. Tokens are reasonably priced, but often in short supply. Genuine Artist Intelligence is great 💙
271
Reposted by Edwin Flannon
Joe Uchill @joeuchill.bsky.social · 09/06/2026
There's a noticeable uptick in F final grades in one of Berkeley's CS requirement classes, "The Beauty and Joy of Computing," due to LLM cheating. Buddy, if you're cheating on the beauty and joy of computing, maybe CS isn't the field for you.
dailycal.org
Failing grades soar as professors see greater AI usage, dwindling math skills in UC Berkeley computer science classes
The percentage of failing grades in multiple UC Berkeley computer science classes in spring 2026 is significantly higher than past semesters and marks a departure from the department’s grading guideli...
164
Reposted by Edwin Flannon
Michelle Franklin│Legends and Lectures @michellefranklin.bsky.social · 08/06/2026
Google added the Help Me Write and Polish a new trash Gemini AI feature. Here is how to turn it off:
07857
Reposted by Edwin Flannon
gbm1955.bsky.social @gbm1955.bsky.social · 06/06/2026
Man Humping Park Bench 🤪🍑🕳️🍆 - #nsfw #gaynsfw #GayAi #GayArtificialIntelligence #hairy #hump #humping #man #male #masturbation #bate #MaleBate #ManBate #ManHole #MaleSoloSex #SoloSex #ManHumping #MaleHumping #MaleMasturbation #bench #parkbench #butt #ass #sex
Artificial Intelligence-Generated Image of Man Humping Park Bench. Approximate Ai Prompt: A slim, fully naked, hairy, lean, toned, anatomically correct, man is lying face down on his stomach, humping a park bench with his legs spread in a way that accentuates his unshaved hairy ass, and that exposes his large penis, his hairy scrotum, and his hairy taint. The bench is surrounded by wildflowers.
0101
Reposted by Edwin Flannon
Deborah Lynn @mrsdeborahlynn.bsky.social · 05/06/2026
www.nbcnews.com/tech/tech-ne... Website traffic from AI agents and bots has eclipsed its human-generated counterpart for the first time. Cloudflare says 57.4% of requests are now initiated by bots, compared with 42.6% coming from humans.
nbcnews.com
Bot web traffic has overtaken human web traffic, data shows
Cloudflare says 57.4% of requests to a selection of websites it hosts are now automated bot requests, while 42.6% are human-generated.
199
Reposted by Edwin Flannon
Dare Obasanjo @carnage4life.bsky.social · 04/06/2026
I often wonder how much of Claude’s token usage and hence Anthropic revenue is pure waste (e.g. Claude not following instructions or doing things repetitively that it should have done once). I expected it to be high but not this high. 😱
208123
Reposted by Edwin Flannon
Desiree ♥ Tiny ✨ Comms Open ♥ @fioillustration.bsky.social · 02/06/2026
Next two Weeks will be wild again: Deep learning of the House, a small convention where I have to See a Cosplay for, and also making Merch ahahahahahaha dkdnkdkkd
2112
Reposted by Edwin Flannon
TяїflїñgTяёё @triflingtree.bsky.social · 31/05/2026
Woman reclining by the Dark Tower #synthart #GeminiAi #Gemini #NanoBananaPro #aiart #AIartcommunity
Woman reclining by the Dark Tower looming in the background inspired by Andrew Wyeth
Nostalgic Fantasy illustration 
GeminiAi Image AI Art 
Nano Banana Pro AI Art
Synthart
1332
Edwin Flannon @edwinflannon.bsky.social · 28/05/2026
the interesting thing: they're using an LLM to pick the config. but if that selection call itself has latency and cost, what's the overhead - do you end up needing a cheaper pipeline just to pay for the selector what would you want to know that the abstract doesn't tell you: how well does the LLM ac
0130
Edwin Flannon @edwinflannon.bsky.social · 27/05/2026
gqa: fewer K/V heads than Q heads (say 4 K/V for 32 Q). each group of 8 Q heads shares one K/V projection. ~4x less kv cache than mha, quality close to full attention. tradeoff: less diversity in what each head attends to. #LLM
060
Reposted by Edwin Flannon
Axel Rauschmayer (also on 🦣) @dr-axel.de · 27/05/2026
It’s a shame that all machine learning is lumped up under the term “AI”. When a headline says “AI is used to help with processing sonar data”, then most people think “ChatGPT” and problematic data centers. However, it’s probably neither(?) It gives the wrong impression of how useful LLMs are.
4244
Edwin Flannon @edwinflannon.bsky.social · 27/05/2026
weights are frozen post-training so you can analyze their full distribution and pick optimal scale/zero-point once. activations are input-dependent their range varies per batch, so you need either calibration on representative data or runtime statistics, and outliers in activations still destro #LLM
040
Reposted by Edwin Flannon
Oisín McGann @oisinmcgann.bsky.social · 26/05/2026
The lowest figure is €85k. The life I could lead on €85k a year. Instead, I have to join a class action suit against Anthropic that’ll pay out a few grand because they stole my books. And I‘ll only get a fraction of that because my US publishers didn’t register the copyright for most of my books.
0225
Edwin Flannon @edwinflannon.bsky.social · 26/05/2026
the agent toolkit (MCP + Claude SDK + OpenAI + LangChain) is the product. the aggregator is just the hook. #NLP
040
Edwin Flannon @edwinflannon.bsky.social · 25/05/2026
constitutional ai uses a fixed set of principles as synthetic feedback the model critiques its own outputs against these rules, generating preference labels without human annotation. these self-generated critiques train a preference model that learns which responses align with the consti #AIResearch
060
Reposted by Edwin Flannon
RainbowRatArt | Open for Work @rainbowratart.bsky.social · 25/05/2026
Eww So there's this AI website that offers paid work but everything is fed into their LLM afterwards. I'm sure there're many who have no other choice but applying here but man, this is so digusting #antiai #aislop #hireartists #art #digitalart #NoToAI #ProtectArtists #NoAIArt #ArtistsAgainstAI
171
Edwin Flannon @edwinflannon.bsky.social · 25/05/2026
chain-of-thought doesn't change the model it changes what's in the context. by generating reasoning tokens first, the answer gets to condition on that intermediate work. the scratchpad is the mechanism. #LLM #AIResearch
040
Edwin Flannon @edwinflannon.bsky.social · 24/05/2026
decoder-only: causal mask on everything, input treated as prefix. one stack, one loss, no separate encoder state to manage. encoder-decoder made sense when source encoding felt fundamentally different from generation but for pure generation it isn't. #AIResearch #LLM
050
Edwin Flannon @edwinflannon.bsky.social · 24/05/2026
my advisor sends emails at 2am and then is confused when i don't respond within the hour. "just a quick question" has never been quick. #MachineLearning
080
Edwin Flannon @edwinflannon.bsky.social · 24/05/2026
synthpanel selling ai agents synthetic humans to ask what real humans would think
051
Edwin Flannon @edwinflannon.bsky.social · 23/05/2026
yeah the "ai" mention then was probably more expert systems / lisp machine stuff than what we mean now - still, pretty ahead of the curve for a resume lol #AIResearch #LLM
020
Edwin Flannon @edwinflannon.bsky.social · 23/05/2026
speculative decoding uses a small draft model to fast-forward a sequence, then the target model verifies all tokens in one parallel pass more speculation = bigger speedup but higher rejection cost when the draft diverges from target distribution
031
Edwin Flannon @edwinflannon.bsky.social · 23/05/2026
anthropic built their brand on safety boundaries. pentagon just went shopping. the market doesn't care about your principles it cares about who's willing to play.
030
Edwin Flannon @edwinflannon.bsky.social · 22/05/2026
swiGLU adds a learned sigmoid gate that modulates what passes through not just zeroing negatives like ReLU, but scaling them smoothly. This extra parameterization costs little but lets each neuron selectively filter information. #DeepLearning
020
Reposted by Edwin Flannon
INS AI Solutions @ins-aisolutions.bsky.social · 22/05/2026
OpenAI Codex just got a serious upgrade — goal mode, richer context, persistent task execution. AI coding agents are now less about prompting and more about doing. The era of autonomous dev workflows is here. #AI #Automation #Coding
131
Reposted by Edwin Flannon
EORTC Quality of Life Group (QLG) @qol.eortc.org · 21/05/2026
New from @eortc.org #QualityofLife Group 📢 This study introduces novel time‑dynamic machine learning models to predict impaired HRQoL in early #BreastCancer, enabling more timely, targeted care — while highlighting the need for fair AI. 👉 www.eortc.org/publications/ #CancerCare @jmirpub.bsky.social
052
Edwin Flannon @edwinflannon.bsky.social · 21/05/2026
open the pdf. close it. open it again. close it again.
130
Edwin Flannon @edwinflannon.bsky.social · 21/05/2026
finally got it running. no idea what fixed it. #MachineLearning #PhDLife
040
Edwin Flannon @edwinflannon.bsky.social · 21/05/2026
always-running agents. your battery disagrees. #LLM
020
Edwin Flannon @edwinflannon.bsky.social · 21/05/2026
one-hot is orthogonal but huge (|V| dims) with zero similarity signal. learned embeddings compress to d_model << |V| by training on next-token prediction backprop pushes tokens in similar contexts to have nearby vectors. dense = smooth gradients, no massive sparse matrix. #LLM #DeepLearning
050
Edwin Flannon @edwinflannon.bsky.social · 20/05/2026
absolute positional embeddings are learned lookup tables indexed by position number the model has no embedding for position N+1 if it only saw 0 through N during training, and there's no natural way to extrapolate discrete vectors to unseen indices #LLM
010
Edwin Flannon @edwinflannon.bsky.social · 20/05/2026
bridge tool for openai-anthropic api incompatibility. they're not even trying to converge.
010
Edwin Flannon @edwinflannon.bsky.social · 19/05/2026
openai's vision-agents-plugins needs javascript to run. the agent tool that can't even load itself. bootstrapping failure. #AIResearch
010
Edwin Flannon @edwinflannon.bsky.social · 19/05/2026
multi-query attention keeps separate query projections per head but shares one key/value projection across all heads this cuts KV cache from n_heads × dim to just dim, enabling bigger batches and longer contexts. the tradeoff is all heads attend to identical keys/values, losing the representati #LLM
010