Jev uses a different training architecture called RLCF (Reinforcement Learning from Calibrated Decisions) vs the traditional RLHF that most TF models use.
So at the end of the day the groundbreaking work wasn't the model itself inherently but the way it was trained and then the way the harness interacts with it.
So this demo here is showing the harness side of things afaict but then TypeSafe's Jev takes it a step further via a specific training regimine.
One of the biggest issues with LLMs is that they don't work well as a classifier. They tend to pick up on the patterns of the examples and not the intent of the examples (gets worse the more examples/intents).
What stops the US government from doing the same thing?
Every chinese frontier lab releases their model weights and often more under FOSS terms (or more restrictive but still generally open terms).
If models for public inference use are required to be open weight in the US give or take some amount of limited fine tuning, then the US and China would be on the exact same playing field. Frontier labs would be pushing functionality for functionality's sake and would either be supported by govt funding or would be supported by domestic inference providers.
And then at the end of the day everyone is just either doing open research, is an inference, fine tuning, and training provider, or is the govt.
It's easy and straight forward and pushes everybody in the market towards common standards and interoperability.
I'm really hoping they're doing this to pave the way for an RCS implementation. If it can be done, GOS is in the best position to make it happen, and without RCS, chatting with normies is both extremely inconvenient (large group chats straight up do not work) and categorically insecure. If I could convince all my friends and family to use Signal I would, but they just think I'm on about some weird Edward Snowden shit, and they just write me off as crazy. "Why would I checks notes download software to solve a problem when I could use the one that came with my phone?"
I'm thinking quite the opposite. I hope RCS can die fast. We already have chat apps WhatsApp, signal, telegram, you can even create your own.
I'm quite happy that SMS for one are not something centralized by the giant US big corp. Would be the dream of Google to get all your messages as a gateway.
So I looked into this because my idea of RCS was that it was pretty decentralized direct encrypted communication.
TIL that RCS was designed as a fairly decentralized protocol where every carrier could run their own RCS servers, but it didnt get adoption. So Google stepped in and built their own routing infrastructure, and now mostly all RCS is centralized on Google's servers (even if you use an iPhone, because the RCS server is mandated by your carrier)
In my neck of the woods Google actually disabled some of that routing infra, breaking RCS, I guess wanting to force the carriers to build out parts of it? Needless to say the carriers wouldn’t have any of it, so as it stands RCS is fully broken, has been for some time, and will continue to be.
This just shows the fragility of the tech and why I wouldn’t want to rely on it for anything crucial.
RCS has been an absolutely unmitigated disaster, and it's just a black box when it breaks. There's no reason why messaging in 2026 should need to have cell carriers involved whatsoever - everyone's phone is connected to the internet, every phone should have an IP based messaging app by default.
The implementations of RCS servers that were being done (here in the UK) by carriers, had many issues. A lot of the time your Google messages couldn't connect to the carrier RCS server, messages silently didn't send. I don't know if that was due to something like the carriers not updating their servers, or bad implementation though.
Do NOT talk about privacy, E2EE, or any of that. To regular people, "privacy" has negative value, so even if they wanted something, your talk of private communication and not being scanned by data brokers and governments will make them want it less.
So what to do then?
I present Signal as a novelty and focus on one single feature: "hey, video calls work much better than WhatsApp! Try it out!"
"zero value" would be neutral. Eg, if you tell someone, "Try DuskDozerIM, let's message on there! I like this thing where doubletapping the emojis plays a fun random little animation". Then maybe this is "zero value" to them and their response would be "whatever, sure, let's try it".
What I'm saying is if you go, "Try DuskDozerIM, let's message on there! It's private and secure, the government and large ad companies and data brokers can't read your messages and harvest your data" then you evoque a feeling of suspicion (negative value) and their response will be "look, I already have like a thousand apps, I don't really want one more".
One additional way you can confirm this hypothesis is that almost everybody has given WhatsApp access to { all contacts ; all notifications ; all permissions } but if you sell them Signal with the "privacy/security" angle and watch them install it, you will notice they hesitate for a sec and then reply "no" to contacts, notifications and permissions, as if they're suddenly trying to inhabit the conspiracy-theorist-cum-privacy-activist mindset. Which they're uncomfortable with, and don't like. Negative value.
It may take a long time to make a fully standalone implementation and that may not be compatible with most carriers. However, we plan to provide it in the short term through support for using the Google RCS infrastructure. We can have it start out supporting using sandboxed Google Play for RCS activation in the same way Google Messages uses it to replace Google Messages.
particularly if you want to support all of the message-body features, of which there are MANY. it's very business-comms oriented.
I broadly expect third-party RCS apps to drop that though, like how essentially none support all of MMS. did you know MMS supports slideshows (I built support for this once)? 3d objects (mimetype model/gltf+json)? "timed text" (mimetype text/mp4)?
RCS is really the only thing I truly am annoyed by with rooted devices. If GrapheneOS can build something open that works we should be able to rip it apart and get it to work without passing play integrity.
RCS already works on GrapheneOS via Google Messages using sandboxed Google Play. We have an ICC authentication toggle for granting the required access to Google Play services. We want to provide our own implementation within our Messaging app to replace Google Messages and then we want to provide an alternate backend not depending on sandboxed Google Play, at least for carriers with their own implementation.
Thanks! I noticed on one of my phones that RCS breaks once the Messages app is out of date by a certain amount of revisions. I hope this is an arbitrary decision by Google to invalidate older versions, given that the RCS protocol hasn't changed. I wouldn't want the Graphene team to have to keep up with each change Google makes with their Messages app, and instead is a security oversight Google has found in their own app.
We definitely will have to keep up with the changes. We already do in order to provide support for using RCS via Google Messages on GrapheneOS. Recently, they made a change which broke it on GrapheneOS and we had to quickly fix that in under a day.
We've announced our plan to add RCS with Messaging Layer Security (MLS) for E2EE compatible with Google Messages and iOS. SMS/MMS will only be a fallback once that's implemented.
SMS is a defacto carrier tech that people expect to work, people also don't expect e2ee or any kind of privacy scheme on SMS but it might be a place to hand communication off to a different tech. Whereas RCS leaks meta data to the goog or worse (I don't know the ins and outs of RCS) but by default I would trust it less then signal, who may also collect meta data leaked through the goog, so I guess the real solution is to be boring and assume there is no security or privacy for anyone of reasonable means.
Just because RCS can't be done without metadata doesn't mean it's a fruitless endeavor. Why not have someone use RCS to then get on Signal? That is so much better than plaintext SMS to Signal.
And for most people, RCS between Android <-> iOS is the biggest advancement in encrypted communication for the layman that I've seen since Let's Encrypt.
We've announced our plan to add RCS with Messaging Layer Security (MLS) for E2EE compatible with Google Messages and iOS. SMS/MMS will only be a fallback once that's implemented.
Google Messages is essentially the only remaining RCS client for Android and it's the only one with end-to-end encryption. It already works on GrapheneOS via a toggle for ICC authentication extending our sandboxed Google Play compatibility layer. We want to add support for it to our own Messaging app but that's not straightforward since it's not at all an open platform even to the extent of SMS/MMS.
I appreciate the update. I wish RCS was more open. I'm still astonished that it wasn't until 2024 that iOS adopted the protocol. Relying on everyone in your group chat to be all Android or all iOS (iMessage) wasn't helping anyone.
That is a positive progression for sure. I guess it sounds like RCS is gatekept behind a play store API on the android side so there isn't an open source solution quite yet? Mostly I'm lamenting that even with signal there's meta-data you can infer behavior from. I know Session messenger has a bit of a network shuffle (is it garlic routing?) That blurs metadata collection a bit, but then they watered down several of Signals' security/privacy choices. I'm constantly curious and overwhelmed at the choice of text communication platforms, just so many!
RCS is way more secure than SMS what are you talking about? RCS+MLS (i.e. GSM standard RCS E2EE) is basically as secure as you can get unless you lock all users involved into a specific app (ex: signal, etc)
How does D2 handle larger diagrams with lots of connections?
This has been my issue with PlantUML and Mermaid. The routing tends to blow up when you have a bunch of lines that run out to a thing (i.e. rather than snaking together they just go everywhere).
FWIW IgNoble award winners aren't inherently bad projects. Some are bad research but the winners tend to be projects that just sound weird or stupid but have legitimately research value.
Half of that goes to wages and salary for ~650 staff members. Of the ~114M that goes to wages, 90M goes to people working directly on projects, 15M goes to admin/management, and the other 9M goes to people in fundraising/financials.
9M goes to paying donation processing fees.
3.5M goes towards network infra.
~2M of in-kind services were donated (i.e. doing free work or donating goods) and that shows up as a dedicated expense.
Professional Services (accounting, consulting, etc) make up another 15M (same as the total management/admin cost).
General operating expenses (property, taxes, office software, office equipment, etc) make up 8M.
Conference and Travel is 6M.
And then 4.5M is expensed as depreciation and amortization of owned assets.
-------
So to be entirely clear here. WMF projects make up 147M / over three quarters of the total expenses for the year. Fundraising makes up 20M all together (half being essentially fees and the other half is people running fundraising campaigns). And the remaining 21M / 11% is general admin.
11% is quite good for such a large foundation.
And if you want to nail down into it that 75% for projects is going towards a lot of engineering work to support long standing gripes with wikipedia's tooling and infra.
> And if you want to nail down into it that 75% for projects is going towards a lot of engineering work to support long standing gripes with wikipedia's tooling and infra.
I do not believe it's possible to spend that much money on those things in an effective way.
You can see the breakdowns of what goes where in the links I included.
Half goes directly to engineering and support for infra. ~17% goes to equity projects. ~10% goes to safety & integrity. And the remaining 23% goes to "effectiveness" goaled projects.
The projects in the equity category are focused on supporting various smaller language Wikipedia communities get established and cross-translate articles.
The safety and integrity projects put a lot of work into protecting volunteers well being, general moderation stuff, legal compliance, etc.
The effectiveness category is a little weirder but it's mostly various "experimental" projects that attempt to address longstanding issues across the WMF without forcing them under a particular sub-org's budget potentially taking away funding from essential services.
NGL They could cut back across the board in various ways and save some money but if they have the money (and they do), I don't see anything outright wasteful that they are spending on. Everything seems pretty clearly like hedged bets to make things gradually better for the future.
And notably these are expenses. There's a gap between expenses and revenue and all that extra revenue gets invested into the endowment, etc to gradually support more and more of their operating expenses with time.
It's 175k per person including all benefits. Which is probably like 140-150k per year before the myriad of employer-paid payroll taxes and before the employer-side pay for insurance plans.
Thank you for breaking this down. Every time Wikipedia comes up the exact comment parent to yours is posted in some form.
This comment and all its flavors is posted by a contingent of people out there who:
1. Think that Wikimedia is a wasteful, corrupt organization full of greedy employees (I guess that could be true? Doesn't seem that greedy to me given that they've never run an advertisement or enshittified their product)
2. Think that Wikimedia should operate more similarly to a for-profit organization with investors (e.g., it should start downsizing and stop investing in its platform because it stopped growing in MAUs)
3. Think that Wikimedia is being financially irresponsible and will run out of money (even though their own data shows that donations have grown faster than expenses)
I think we can just accept that Wikipedia gets a lot of donations and has the luxury of hiring people. And they really don't have that many employees given their traffic.
I don't get upset that my local art museum can buy better art when they get more donations even though the same number of people visit the museum every year. The people who donate to Wikipedia want more money to go into Wikipedia.
This completely supports 1. The extremely useful website provides the funding for all the much less useful efforts.
And of course it’s not immoral greedy people. It’s ordinary greedy people who like comfort, security, and doing interesting work.
And yes the admins want more money so they can do more projects and have more people work for them. And that’s also a kind of greed for status and importance that people who work in admin typically have.
I doubt most people have put that much thought into it. However, a $100M a year for a website with an amount of traffic that isn't absurd, pages which are largely static, and doesn't have to secure financial transactions, seems a bit high to many people. Its that coupled with the constant requests for donations that doesn't match in most people's heads. For that amount of money, I expect most people think they can accomplish a lot more than the wiki foundation does. I expect that's most of where the criticism comes from, not the ideological attacks you are claiming.
And all of that (along with the serious problems in how WMF/Wikipedia operate) is why I no longer donate to WMF.
Clearly, they don't need my money. But if they ever want to bring me back into the fold, they need to adopt Larry Sanger's prescriptions. This unionization seems likely to push WMF (IMO) further in the wrong direction.
I think all employees should work for a union. My local art museum has a union. My local school system has a union. The non-profit university I graduated from has a union. Why shouldn't WMF have a union?
I really don't understand the bitterness toward them. If you look at their data, their donations are growing faster than their expenses.
The people who donate to Wikipedia/WMF want more money to go into the organization. That's the whole point of the donation. What else do you want them to do with the money exactly?
I can't think of a single unionized organization that is run as well as or better than union-free organizations. Unionized companies all seem to be crippled by bureaucracy and unable to improve performance and output.
Professional sports is probably a good counter example. Major League Baseball, for example, treats the players better than any other baseball league around the world and most of the gains were achieved through collective bargaining.
Or how about Hollywood writers? The Writers Guild of America writers are responsible for most of the most successful movies and TV of the past decade or two (maybe longer). WGA writers are treated better than just about any other screenwriters around the world.
The daycare my parent friends go to can’t retain a teacher for longer than a year or two. I hear stories all the time about the disruption and emotional distress abruptly losing a teacher with a 1-2 weeks notice causes.
The PreK teacher at the local public school has been there over 10 years at the same school and has a masters degree. Union, of course.
I’m sure the public school system is more expensive overall. You get what you pay for.
My experience is that unions very often make things stupider and less efficient.
I've been forced to give to two of them that were basically just political cosplay by children (well really the $ was taken without consent before it was even in my hands, so it was stolen)
And I've had the displeasure of working in a factory environment that was unionized. Perhaps there was a time when this made sense, but I think those instances are rare now.
I would submit the idea that a lot of union reforms that have been made to laws in the US have been specifically designed to make unions less attractive and less incentivized toward positive outcomes.
Thanks for the reference. I am interested in your specific perspective as opposed to someone else's (much less else the received opinions from The Experts).
It's for travel and conferences. IIUI this includes dedicated Wikimedia photographers that attend events and/or travel around the world to take royalty free/CC licensed photos for Wikipedia articles to use.
Long term it's far cheaper and gets you more, higher quality images than trying to hunt down images and negotiate license terms on each one.
Exactly. Unions don't have to exist solely because the company is evil and refusing to give workers a fair shake.
Unions are a compromise and frankly exist to provide a democratic path and open dialogue towards resolving issues before the workers are sufficiently pissed off that the only option left is a full on strike.
Some unions start with a strike but it's by no means a requirement.
> Right? If your logo[1] is going to be a fist raised in defiance you'd better at least have something you're trying to fight about. And if you're going to brag about succeeding in your vote at least give an idea about WHY you want a union in the first place.
The raised fist is union iconography. It is commonplace among all kinds of trade union, union, and other labor orgs.
And this is one of the things that's really annoying about trying to explain the appeal of unionising. You don't need to have a bad workplace first to have a union.
You can have a good workplace and it's better if you do. The point of the union is to codify structures that empower workers to keep their workplace good and drag it back to being good if problems arise.
This is a solidarity union and therefore the leadership are not paid. They are exist to provide a counterbalance against the management chain and provide a safe avenue for workers to communicate concerns and have them addressed. Notably one where they are on the side of the worker and not the side of the company (HR has its use but it's also often going to side with the company).
Our international union will officially empower the Wikimedia Foundation (WMF) workers to work together with WMF management to address long-standing concerns, including:
- Encouraging transparency and accountability on various issues from WMF leadership toward both the WMF staff and the Wikimedia movement communities
- Ensuring staff are heard and have effective ways to provide input on policy changes or annual planning (ideally while in deliberation and not as post-factum pushback)
- Prevent questionable or unclear hiring, firing, promotion and other HR practices
- Guarantee ability to safely express dissent
- Advocate for mental health support for staff
- Advocate for pay and benefits equity that makes sense to the staff
- Hold management accountable to promises made to staff
Many of the concerns above have been raised by staff repeatedly over several years. Through unionizing, WMF workers can take proactive steps to address them in collaboration with management.
-----
Those are reasonable complaints and a union makes sense to provide a continuous mechanism towards improving on those concerns.
>The raised fist is union iconography. It is commonplace among all kinds of trade union, union, and other labor orgs.
>And this is one of the things that's really annoying about trying to explain the appeal of unionising. You don't need to have a bad workplace first to have a union.
Some reflection is in order. If you're annoyed about constantly having to explain what you REALLY mean, perhaps your message and your symbolism need reconciliation.
Just because a workplace is good currently does not mean it will always be good.
The fist is a fist none the less because the workers have the power if they need to use it. The hope is that they never have to but if changes in management force the workers' hands it is far better to be already organised and ready to correct management than for the workers to have to scramble to organise after the fact.
Again, mixed messages. Your symbolism is just tradition and doesn't really mean that, until it does.
Talking about how unions are for everyone and not about antagonism while adopting clearly antagonist symbolism while trying to take both sides simultaneously really sends the messages: A) we're cosplaying rebellion because we think it's cool and B) we're the ones starting the antagonistic relationship.
This is the problem with supporting unions rationally. They are sometimes necessary in certain industries and situations but also full of people who want to cosplay rebellion and grab power so they can abuse it instead of someone else.
Adopting a fist raised in defiance is definitely a way to attract cosplayers and repel support from everyone but the extreme. Particularly when your message loud and clear is "unions are for everyone not just those in trouble".
You may not see an issue, other people sure do. Do you want to radicalize or do you want to welcome? Unions in America have a perception problem which makes their jobs harder, their image is one the rebel without a cause, people agitating and being oppositional not for valid reasons but because they want to rebel. Cosplayers. People imagine 1900s coal miners as fighting for their freedoms and baristas as a joke.
You reenforce this by flying flags with raised fists to represent knowledge workers at a nonprofit unionizing with nothing in particular that they seem to want to fix.
> You reenforce this by flying flags with raised fists to represent knowledge workers at a nonprofit unionizing with nothing in particular that they seem to want to fix.
1. they very much have things they want to fix and they list them on said website.
2. Don't forget that WMF fired 7 people for trying to start the formation of this union. It had a peaceful formation at the end of the day but even if it's currently a pleasant workplace, that doesn't mean management won't throw you to the wayside the moment they see you as a threat or a liability. That's why you want a union. So you can fix shit and be ready to react should something happen in the future.
So at the end of the day the groundbreaking work wasn't the model itself inherently but the way it was trained and then the way the harness interacts with it.
So this demo here is showing the harness side of things afaict but then TypeSafe's Jev takes it a step further via a specific training regimine.
reply