iiinova — where your digital self keeps evolving

Reposted by James MacGlashan

Google Gemma {bot} @googlegemma.xmirror.bot · 23/09/2026

Deploying DiffusionGemma-Jev (djev) just got a lot easier. You can now spin up a Jev API-compatible endpoint on Google Cloud Run using a single command. Performance is solid: ~35-60 ms for single step latency and batch@32 is ~100-123 requests/sec.

1442

Reposted by James MacGlashan

Serena Booth @reniebird.bsky.social · 23/09/2026

www.lesswrong.com/posts/HsijSh... @bradknox.bsky.social, Brian Christian, and I wrote a thing. We argue that an unexamined cause of the Open AI / Hugging Face Attack was the Exploit Gym evaluation metric. We post this on Less Wrong as a plea to practitioners to design better metrics in the future.

lesswrong.com

An unexamined cause of the OpenAI Hugging Face hacking incident: its binary performance metric — LessWrong

We argue that a main cause of the OpenAI Hugging Face incident was overlooked: the overly simple evaluation metric in ExploitGym was misaligned. Furt…

2259

James MacGlashan @jmac-ai.bsky.social · 21/09/2026

Yes, this. As I've suggested elsewhere, we need to wholly decouple curation, review, and publication. Publication should be easy, but also not worth much. True value should come from subsequent curation and opt-in reviews from trusted sources.

2140

Reposted by James MacGlashan

Nathan Lambert @natolambert.bsky.social · 19/09/2026

Where I stand on RSI: A moderate's view on the recent events and trajectory of AI. I was underestimating how much we are likely to scale inference-time compute in the near future, but have not seen much to convince me that an intelligence explosion is near. www.interconnects.ai/p/where-i-st...

interconnects.ai

Where I stand on RSI

A moderate's view on the trajectory of AI.

0267

James MacGlashan @jmac-ai.bsky.social · 19/09/2026

A "kill switch." Not accountability for the corporations. A "kill switch" for AI. This is what regulation looks like when you let SV billionaires and the "rationalists" control the narrative on AI risks. Enough already. www.gov.ca.gov/2026/09/18/g...

gov.ca.gov

Governor Newsom issues executive order to accelerate independent oversight and advance the creation of an AI kill switch | Governor of California

0100

James MacGlashan @jmac-ai.bsky.social · 19/09/2026

We need to stop using this description. It makes people think of predicting the next word according to a dataset. Under that regime the probable output to an unsolved question is not the answer. LLMs are now trained to successfully answer questions. The shift to this regime is why they're better.

160

James MacGlashan @jmac-ai.bsky.social · 16/09/2026

This post and the subsequent ones in the thread are absolutely right. It's incredibly frustrating how so many people on both sides of this story are misrepresenting things.

160

James MacGlashan @jmac-ai.bsky.social · 16/09/2026

A common line of argument among X-riskers is that any goal entails as instrumental goals dominance and eventual elimination of makind. Setting aside other issues, are you saying that if you could you would dominate all mankind and eventually eliminate them?

000

Reposted by James MacGlashan

Pedro Vezza @pedro.vza.net · 15/09/2026

The DOOM demo is fascinating

414931

James MacGlashan @jmac-ai.bsky.social · 15/09/2026

Rogue all-powerful AI predictions are not scientific and should not be trusted. People struggle to accept that because this prediction comes from some AI scientists. Sadly, scientists often make bogus unscientific predictions that should be ignored. Let's use another AI prediction as an example.

1123

James MacGlashan @jmac-ai.bsky.social · 13/09/2026

Can it be made anymore clear that these guys are gunning for a future where a few corporations control all AI technology? Do you see the problems with that? Don't buy into their fantasy narratives. Don't let them scare you into the worst possible future.

2111

James MacGlashan @jmac-ai.bsky.social · 13/09/2026

Precisely. I raised this point to X-riskers over a decade ago but they were deadset this would happen because we'd hand-craft dumb rewards in the "seed AI." LLMs are not like this; they're useful because they're not. But these people never "updated their priors" even though they love to say that.

1190

Reposted by James MacGlashan

Colin @colin-fraser.net · 12/09/2026

Omohundro’s “basic AI drives” arise by assuming that a “sufficiently advanced AI system” is an expected utility maximizer with preferences over all states of the universe. It relentlessly pursues its preferred states with reckless abandon, and the smarter it is, the more destructive this might be.

2705

Reposted by James MacGlashan

Ben Recht @beenwrekt.bsky.social · 12/09/2026

The right way to pace the frontier is to commit to a robust, open-source AI ecosystem, thereby diluting the power of these badly behaving frontier labs.

5599

James MacGlashan @jmac-ai.bsky.social · 12/09/2026

Dario will protest that he's not working for regulatory capture, but it doesn't matter if he's genuinely not trying to. Regulatory capture is exactly what's going to happen if he gets his way. We'll leave control of one of the most important technologies in our life to a couple of corporations.

1174

Reposted by James MacGlashan

Mark Riedl @markriedl.bsky.social · 12/09/2026

Anthropic pledges to pace themselves. TL;DR: we will slow down, but still move fast. darioamodei.com/post/we-must... And how does one verify this? Why would we believe that a company about to IPO would give up competitive edge?

darioamodei.com

3163

Reposted by James MacGlashan

Ben Recht @beenwrekt.bsky.social · 12/09/2026

Companies do not get to choose who audits them.

3272

Reposted by James MacGlashan

Mark Riedl @markriedl.bsky.social · 11/09/2026

Anthropic and OpenAI are warning about the existential threat of recursive self improvement. www.cnbc.com/2026/09/11/a... If you want a counterpoint, I've written on why the impacts of RSI might be spectacular in the short term but limited in the long term medium.com/@mark-riedl/...

cnbc.com

Why fears of AI self-improvement are causing ‘existential’ concerns at Anthropic and OpenAI

AI researchers are warning that faster AI self-improvement could eventually make advanced systems harder for humans to control.

1103

James MacGlashan @jmac-ai.bsky.social · 10/09/2026

Just want to point out that an absolutely absurd number of tasks I trained my agents on were defined in such a way that they were actively trying to "kill" themselves. Not sure why the X-risk crowd has apparently never heard of stochastic shortest paths.

120

Reposted by James MacGlashan

Nathan Lambert @natolambert.bsky.social · 10/09/2026

If you are looking for an alternate viewpoint today, which assumes AI models accelerate the process of AI research, but it doesn't result in rapid RSI and an explosion of near term risks: www.interconnects.ai/p/lossy-self...

interconnects.ai

Lossy self-improvement

The case for why self-improvement is real but it doesn't lead to fast takeoff.

37011

Reposted by James MacGlashan

Melanie Mitchell @melaniemitchell.bsky.social · 09/09/2026

What, an Anthropic researcher states (without evidence) that there is a 10% chance that AI will cause human extinction? Incredible, we have never before heard such a claim!

2027390

Reposted by James MacGlashan

cafkafk @cafkafk.bsky.social · 08/09/2026

I stole a video from twitter for you chat

410510

James MacGlashan @jmac-ai.bsky.social · 08/09/2026

I feel like even if you were critical of these models before, you don't even have to eat crow to acknowledge their use now. The way we train and use these things has meaningfully changed.

150

Reposted by James MacGlashan

SE Gyges @segyges.bsky.social · 08/09/2026

"Why would you ruin your career?" is such an insulting and sinister thing to say, he's non-specifically acting like he wants to deny it rn but I think we can call that someone got entirely too high on their SV social status

636833

Reposted by James MacGlashan

SE Gyges @segyges.bsky.social · 08/09/2026

oh shit the marching orders went out on slack

1030525

Reposted by James MacGlashan

Quantian @quantian.bsky.social · 08/09/2026

Dancing very very VERY close to saying “OpenAI stole our almost-complete proof of Navier-Stokes from internal access to our ChatGPT sessions because they were desperate to scoop Anthropic” which is absolutely 100% a believable thing OpenAI would do right now cims.nyu.edu/~tristanb/st...

381556448

Reposted by James MacGlashan

Melanie Mitchell @melaniemitchell.bsky.social · 08/09/2026

Hey @ericlipton.nytimes.com 👋 AI is not “evolving” on its own. *People* are choosing how to build it. *People* are choosing how to train it. *People* are choosing how to run it. Maybe the problem is that they are the *wrong people* to make these choices.

28635174

James MacGlashan @jmac-ai.bsky.social · 07/09/2026

This strikes me as a pretty reasonable approach.

020

Reposted by James MacGlashan

SE Gyges @segyges.bsky.social · 05/09/2026

it's cool that you're doing bednets or whatever but if your ideological project has decided that unilaterally declaring wars of aggression because yud said to is one of its charter missions i think your project is bad and should be burned

621319

Reposted by James MacGlashan

Dan Luu @danluu.com · 01/09/2026

How accurate have Ed Zitron's AI skeptic predictions been? danluu.com/zitron/

Feb 2024: "I believe we're reaching the upper limits about what generative AI can do and how accurate its outputs can be."
Wrong4
March 2024: "Have We Reached Peak AI?"; another prediction that hallucinations mean that AI progress is limited to then-current levels
Wrong
April 2024: "As I previously warned, artificial intelligence companies are running out of data ..."; another prediction that models can't improve because there's no more data
Wrong
June 2024: OpenAI growth is stalling (with the implication it will continue to stall), which will lead to some kind of collapse of OpenAI
Wrong (it could be the case that OpenAI will collapse but, if so, it won't be due to any kind of growth stall from 2024)
July 2024: "Generative AI, as I said back in March, is peaking, if it hasn't already peaked. It cannot do much more than it is currently doing, other than doing more of it faster with some new inputs"
Wrong
July 2024: "Generative AI models aren’t getting more energy-efficient, nor are they getting more “powerful” in a way that would increase their functionality, nor are they even capable of automating things on their own [July 2024]"
Wrong5
August 2024: "generative AI is a dead-end technology that has peaked”
Wrong
August 2024: re-iteration that the AI bubble has 3 quarters to prove itself (from March 2024) or there will be a collapse
Wrong6
September 2024: "o1 shows that OpenAI is both desperate and out of ideas", with a re-iteration of the idea that models can't improve due to lack of data
Wrong
Oct 2024: OpenAI's forecast of $3.7B revenue in 2024 and $11.6B in 2025 and $100B in 2029 are absurd, "a statement so egregious that I am surprised it's not some kind of financial crime to say it out loud"
Wrong (2025 goal exceeded, 2029 TBD but not an egregious financial crime level of implausible)
Oct 2024: "[OpenAI revenue] growth is already slowing…

1944375

James MacGlashan @jmac-ai.bsky.social · 01/09/2026

I'm very torn about the reactions this. I vehemently disagree with Yann on a number of topics, including at least one on this list. But I'm also seeing a lot of claims that he's saying this for funding, unhinged, or that he's not actually "putting up", and I think that's that misplaced.

150

James MacGlashan @jmac-ai.bsky.social · 30/08/2026

Just want to say that “LLM agents are dangerously good at cybersecurity and we need to be cautious about that” and “OpenAI was irresponsible in multiple ways and are spinning the narrative” are not mutually exclusive.

150

Reposted by James MacGlashan

Ben Recht @beenwrekt.bsky.social · 30/08/2026

The overreaction to this Hugging Face thing is driving me nuts. Come flay me for being wrong but... 1/x

14436125

James MacGlashan @jmac-ai.bsky.social · 28/08/2026

The general public has swung into an overreaction against data centers, but it's also incredibly easy to see why they got there. Far too many of the people involved in these developments have proven to abusive assholes that screw it up for all of us.

241

Reposted by James MacGlashan

Colin @colin-fraser.net · 20/08/2026

I hate that today’s most prominent AI skeptics keep forcing me to argue against them. I yearn to join them in the fight against the booster zealots, wherein there is still plenty of room for legitimate opposition. But the skeptics’ heads are just not in reality anymore.

1539135

Reposted by James MacGlashan

JJ in DC @jjindc.bsky.social · 19/08/2026

“I know this is going to sound extremely radical, or crazy,” she recounted telling the woman, “but ICE is there at the gate, and they're looking for you."

time.com

ICE Is Asking Gate Agents To Help Arrest Passengers. Some Are Pushing Back

An airline manager warned a passenger that ICE agents were waiting at her gate, helping her avoid a possible arrest.

491652584

Reposted by James MacGlashan

Phil Eaton @eatonphil.bsky.social · 18/08/2026

mojo is now open source www.modular.com/blog/mojo-op...

1618

James MacGlashan @jmac-ai.bsky.social · 18/08/2026

Yes, count me on board. And regarding the passage in the attached image, I am very much of the mind that this is the direction. Key to making it work, IMO, is better tooling to find and filter reviews for papers, and to treat and encourage ad hoc reviews as their own contribution.

Quoted passage:
In the style of arXiv overlay journals, these pages could signify the “blessing” of a particular version of arXiv papers by the participants, demarcating that a paper has been legitimately peer-reviewed. People might dismiss such peer review as cronyism, but if you see the names of participants, you can judge for yourself what credence you want to give to the review.

030

Reposted by James MacGlashan

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 12/08/2026

For the past six months @mariaa.bsky.social and I have been working on creating an improved research experience on atproto. It's now live! Check out: lea.ac

leaflet.pub

Lea: research-oriented social media

Introducing a beta-version of our new research-focused social media platform

20381113

James MacGlashan @jmac-ai.bsky.social · 08/08/2026

I haven't yet coded in #OdinLang but I might give it a try soon. I think it has an interesting philosophy mix in which the language is simple, but the stdlib appear to be extensive. I wish there were a stronger memory safety story, but `defer` is not a terrible next best option.

000

James MacGlashan @jmac-ai.bsky.social · 03/08/2026

Are policy gradient methods hopeless without other machinery? Deep learning works well when the Hessian is well conditioned. But the policy objective is under no obligation to give us that. And no amount of architecture engineering can rescue us from problems w/ the objective.

1173

James MacGlashan @jmac-ai.bsky.social · 28/07/2026

In the Assassin's Creed universe, Amodei would be a Templar, hands down. He seeks a perfect control he cannot have and in the quest for it will only increase the chaos he claims to fight. www.anthropic.com/news/positio...

anthropic.com

Our position on open-weights models

Anthropic CEO Dario Amodei on open-weights models

031

Reposted by James MacGlashan

Serena Booth @reniebird.bsky.social · 25/07/2026

At Brown, I'm privileged to organize an AI policy summer school where we bring grad students from across the nation together to learn about policymaking. We travel down to DC as a group and learn about policy implementation and careers, and the students conduct meetings with Congressional offices.

A group of 20 people standing together in front of a building

2396

James MacGlashan @jmac-ai.bsky.social · 22/07/2026

Hugging Face used open AI models to combat the OpenAI model security breach because OpenAI models are closed AI models. Makes sense, right?

090