today on the ai daily brief why clawed tag and approaches like it might change how you use ai
before that in the headlines an anthropic customer has sued the us over losing fable
but can it possibly help the ai daily brief is a daily podcast and video about the most
important news and discussions in ai all right friends quick announcements before we dive in
i am continuing to remind folks about the executive agent leadership program which is the
enterprise grade descendant of enterprise claw now offered in conjunction with super intelligent
you can find out about that at training.bsuper.ai the next cohort begins on monday june 29th so if
you're interested in that check it out well here we are heading towards the back half of this week
with a return of fable 5 no closer insight in fact the rumor mill suggests that we are not in line for
pretty much any models for some time later this week i'm going to do a show about what you can
do to make the most of this forced ai model pause but for now let's look at some of the conversations
that are actually shaping the discourse now first up i flagged this before but i want to take this
chance to flag it again right now emotions are running at an 11 when it comes to this anthropic news
i saw a post go viral on x about the significance of this bipartisan letter asking our lutnik if
anthropic had been singled out but friends this is at this point about a week old and these sort of
open congressional letters don't actually do anything they don't usually get responses this is
in a word a nothing burger one thing that also probably amounts to nothing but is at least a little
bit more interesting is that one anthropic customer is taking it upon themselves to get the model
reinstated by suing the government in a lawsuit filed on tuesday a legal tech firm called legion
claimed the ban was illegal and damaging their business the harm to legion is immediate
irreparable and existential the lawsuit stated the pace of frontier ai advancement is blistering and
competitive ground lost during a suspension cannot be regained after the fact certainly regardless of
the ultimate success of this lawsuit i think that is a sentiment that many of you listening now will
agree with the filing claimed that export control laws do not cover shutting down access to cloud
software a point that many other legal commentators made as soon as the ban was announced serving
computer code or text outputs from a computer isn't typically seen as an export that can be proper
subject of export controls in the alternative if the ban relied on ieepa powers similar to sanctions
the lawsuit notes that these powers do not apply to information and no prerequisite national security
emergency was declared further the lawsuit argues the move was overbroad and possibly retaliatory
towards anthropic it also notes the decision to block distribution of a model directly contradicts
the executive order sign at the beginning of june which ruled out the idea of a model licensing
scheme in making their argument legion centers the facts on their canadian development team losing
access to the model pointing out that targeting allied nations with export controls is highly
unusual anthropic for their part remains silent on the matter and there is certainly no indication
this lawsuit is being filed on their behalf by a third party in comments to the press they referred
to previous statements indicating they're grateful to the administration for working quickly to resolve
the issue now in terms of the potential success of this at this point most legal analysts that i've
seen seem to think that the administration has gone outside the bounds of the law on this one
but also that it's unlikely to be a clear-cut decision for a judge and frankly none of that is going to
actually matter because as long as anthropic is taking it seriously which they clearly are presumably
and oh my gosh hopefully it'll be resolved through negotiation before any legal process could actually
make it through the courts now the administration's friskiness when it comes to controlling the flow
of ai is doing nothing but increasing the most recent news is that the administration is now
pressuring meta to agree to voluntary model review sources told the new york times that the white
house is pressing meta to submit their ai models for safety and performance testing prior to release
meta is currently the lone holdout among major ai labs after microsoft xai and google signed
agreements shortly before the recent executive order on ai and cyber security now reportedly the
administration's requests seem to only relate to testing at the center for ai standards and
innovation at the commerce department not whatever additional security testing has been going on at
the nsa and meta does seem to be leaning towards an agreement with a spokesperson commenting
we share the administration's goal of advancing us leadership on robust and secure frontier ai
while we are working through the details we hope to sign the agreement soon still the whole idea of
pressuring a company to submit their technology for government testing really highlights just how
air quotes voluntary the program actually is and having contained advanced ai models the commerce
department is reportedly turning to robots as their next technology to ban politico reports that
commerce secretary howard lutnik held a closed door meeting on monday to discuss the risk of chinese-made
models according to notes provided to politico lutnik said
we don't want state subsidized robotics attacking us in america this is the arms race that is coming
robotic arms are coming we need to make sure they're produced in america so we're going to study
those right now commerce department officials reportedly view state-backed chinese robot companies
as a massive threat fearing they could proliferate through global markets before u.s alternatives hit scale
chinese robots are already subject to steep tariffs but the commerce department is considering
additional action the monday meeting included more than a dozen executives from major firms including
boston dynamics spacex siemens and goldman sachs the major topic of the meeting was how the u.s can
best rebuild its industrial base but apparently the risks of chinese robots dominated the discussion
one source said the whole idea that what we're going to end up with is an american brain with a
chinese body is a very very bad strategic plan now it's unclear from the reporting how fast we should
expect to crack down but the sabers are already rattling elsewhere in washington on monday the x account for
the congressional select committee on china posted a screenshot of a unitary humanoid robot for sale
through amazon it's priced at seventeen thousand nine hundred ninety dollars and can ship to washington
in a month the select committee added unitary was recently designated as a chinese military company
and its products are a threat to our national security yet here is amazon selling a unitary robot
in america we need chairman mulnar's guard act to stop this threat and support american robotics
now several chinese robotics firms have received the same designation which many see as a precursor to
a huawei style ban now on the topic of american-made robots elon musk recently said that they are in the
final stages of completion of optimus 3 which he says is really going to be by far the most advanced
robot in the world nothing's even close but people are i would say fairly skeptical of those claims
speaking of elon though xai has followed suit to openai and anthropic introducing the goals primitive
in an update to their coding harness grok build as with the goal command in codex this allows users to
define an outcome and set grok to work on a long horizon task grok will continue working across
multiple steps deploying sub-agents to complete individual tasks and orchestrating the entire
process without user intervention the orchestration process includes planning implementation and
verification to ensure a high quality output so a couple things to make this worthy of inclusion
here first of all it reinforces the idea that slash goal is an actual new ai ux primitive not just a
generic feature but second this is our first glimpse of what the xai cursor partnership will look like
after the acquisition was finalized last week the goal command natively uses a fine tune of grok
called grok build 0.1 as well as cursor's composer 2.5 to complete its tasks and the combination has at
least the potential to be more than the sum of its parts composer is known for cost-effective high
performance making it a natural fit for coding sub-agents and while the latest version of the
flagship grok model didn't grab that much attention on release one big point of differentiation it tried to
make was a sophisticated orchestration process built into the model when used as a chatbot this mostly
just resulted in sub-agents arguing over how to interpret information but that same process applied
to coding tasks could make grok really good at that sort of orchestration we'll see what comes next and
whether this makes any waves but certainly the update to build reinforces the idea that xai at this
point as much as spacex is becoming a neocloud isn't quite done releasing ai products just yet
speaking of ai products one that is getting a lot of attention is byte dance's seed dance 2.5 we've
started to get a ton of previews of the new model which doubles maximum clip length from 15 to 30 seconds
and adds 4k resolution the new version also supports up to 50 input references to use as objects or
characters within scenes by comparison seed dance 2.0 could handle 12 references and google's vo 3.1
released back in january maxed out at 3. c dance 2.5 is also able to make use of image video and audio
files as references the first time anyone has moved beyond images now so far it doesn't look
like anyone has actually got their hands on the model so we're just going off of a demo clip
provided by byte dance yet that's enough for people to suggest that this will be the mythos
moment of video models simon smith who i'm fairly sure isn't getting paid by byte dance by the way
that's a joke if you've heard simon smith quoted here he was very very much not getting paid by byte
dance wondered how they managed to gain such a big lead in video he asked seed dance seemed to be
pulling away in the video generation race is there a recursive feedback loop at work here if so what
is it are they generating training data with each generation to train the next if so what do they do
to increase quality with each iteration whatever it is the corpus of training content they have access
to through tick tock seems to be paying off lastly today i don't want to get into it deeply but one
story that i am keeping an eye on is some of the market moves that i had previously discussed with
regard to google specifically moving to a broader ai sell-off to the extent that this becomes a more
significant narrative trend as opposed to just a particular market moment we'll come back and
explore it more for now though that's going to do it for today's headlines next up the main episode
welcome back to the ai daily brief sometimes big stories in ai are extremely obvious for example the
u.s government sending a note telling anthropic that they're using export controls to deny access to the
latest model to everyone who's not american and anthropic responding by taking down the model that they
had released just a few days ago for everyone that's a big story in a very obvious way today's
story is big not because of some geopolitical implications but because of the way that it
potentially changes at somewhat of a core level how we think about interacting with ai at work now
the feature we're discussing is called claude tag which claude announced on tuesday as a new way for
teams to work with claude and to put it really simply it's basically claude in slack but not just as a
different channel to prompt claude but instead as a full team member that has access to all of the
context from all the different channels as well as the different tools that your team is already using
anthropic writes claude tag is an evolution of claude code made more proactive and built to work with a
full team it's now one of the main ways we get things done at anthropic 65 of our product team's code
now comes from our internal version tag claude with a request and it breaks the task into stages works
through them with the tools it has access to and responds in the thread with what it creates it'll
write or merge pull requests run data analysis or help resolve an incident in a channel that's a
channel on slack there's one claude that interacts with everyone so a teammate can pick up exactly where
you left off it builds more context about the work as it follows the channel so you don't need to
explain things from scratch turn ambient behavior on and claude takes initiative it follows up on threads
that have gone quiet and flags what's relevant from across its channels and tools so for someone who's
moving fast you could be forgiven for thinking it's just another agent in slack sort of announcement
of which we've had a fair few but just based on the way that the anthropic team was talking about
it there's clearly something going on here anthropic team member alexander bricken writes
claude tag has completely changed the way we do work at anthropic having the multiplayer experience
of claude within slack with all of the fundamental primitives of claude code like proactivity
subagents and long horizon task management is an entirely new way to do work asynchronously
tobin south writes claude tag is how i do 90 plus of my work i even had people from anthropic texting
me specifically when i didn't cover it on yesterday's show basically arguing behind the scenes that this
is a big deal now i had already begun planning this episode when those texts started coming in
but still that should tell you how the team behind it views its significance now this is not the only
feature of this type that's available now anthropic is not the first company to release something like
this the semi-analysis team wrote perplexity's computer slack co-worker has been wildly useful
internally at semi-analysis which is why anthropic launched a computing product the team is configuring
claude tag ai co-workers and is excited to compare and contrast them with perplexity computer click health
simon smith wrote immediately asked our cto to activate claude tag we already use chat gpt workspace
agents in slack they're great for a bunch of use cases like providing support based on prior resolve
slack posts and summarizing slack activity to create a weekly brief the difference with claude tag is
that this level of configuration isn't even necessary you just drop claude into a slack
channel and it picks up the context and becomes useful we currently create one slack channel per
project we've historically also created chat gpt projects and notebook lm notebooks along with those
slack channels to work with ai duplicating or triplicating context across all those surfaces
now we can just create the slack channel as we've always done and drop claude in i also think this will
massively help with ai diffusion people don't have to install claude desktop to access claude's most
powerful capabilities in co-work and code they can just interact with claude as if it's another
person in slack no less than andre carpathy argued that this is indeed a bigger deal than it seems at first
he wrote this is a new paradigm for interacting with claude that is significantly more in line with
all the other human activity org wide once you do all of the under the hood engineering work to make
this just work eg across tools integrations compute environments memory security etc claude basically
joins the team in a seamless way you can talk to it as you would talk to a person and it can help with
a very large variety of workloads in my opinion this is the third major redesign of llm ui ux the first
paradigm was that the llm is a website you go to the second was that it is an app you download on your
computer the third one is that it is a self-contained persistent asynchronous entity with org wide tools and
context working alongside teams of humans it really takes a while to wrap your head around it but it
works and it is awesome now for a lot of folks the big eye blinking statistic was that anthropic is
saying that 65 of their product team's code now comes from claude tag ejaz wrote it's blowing my mind
that 65 of product code in anthropic is now written by tagging claude in group chats of staff discussing
what they want to build r.i.p the days of tediously writing long product docs now you can literally go
from slack to a production ready feature cloud code is barely a year old by the way developer nick
dobos wrote anthropic isn't using cloud code anymore major agent ux shift alert and i think this is one
of the sneaky things that's going to take people a little while to fully wrap their heads around
while you might be tagging at claude what you're really calling upon is claude code now certainly
both anthropic and openai have already spent lots of cycles trying to make the coding capabilities of
their tools more accessible for non-coders but this is the biggest move in that direction yet
when claude already exists in the workspace that you're operating in the barrier to using it to
build something is having an idea and dropping it into the slack where you're already working
now while it's early days for non-anthropic users many were quick to share what they had learned from
using claude tag over the past few weeks or months claude code team member tarik pointed out that each claude in
each channel is different one suggestion he had was to introduce air quotes claude to a new channel
with a pin message and a set of instructions to remember like which persona to take when to respond
etc he says you can think of it like it's claude.md because all of this will get added to its memory
tarik also suggests having a personal channel eg tarik-claude where you can tag claude for your own
work and give it instructions specific to you you can then forward other messages to that channel so your
claude can get to work the way you like he also noted that this helps with overwhelm because when things
get going it can be hard to keep track of all the threads he wrote in my personal channel i like to
ask claude to keep a pin message that it updates with the status of everything he even uses a set
of emojis to be able to help him understand where different work streams are in a single glance
chris taylor from fractional who joined the broader anthropic ecosystem to run their version of a
forward deployed engineering group described how they had been using it across their company he said
it kind of works like a co-worker that just picks things up it owns busy work daily status
summaries following up when people are out of office nudging teammates on open to
dues pushing back due dates his plans shift chris said that when one of its own scheduled tasks
started misfiring it caught the bug fixed it and wrote it up without being asked he also said that
it manages itself we needed somewhere to track bugs and ideas for claude tag while we were testing it
so it built that tracker for us and files every new issue and use case on its own finally reinforcing
that this is in fact a fully fledged version of claude code not just claude he says it writes and ships
handed one of our code bases claude read through it wrote up the 10 most important problems to fix
stack ranked then made the changes and opened them for review all in slack the claude dev team shared
that they're using it for incident response bug triage dependent work i.e work that's blocked on
something else background watching and monitoring metrics nityesh from the every team wrote about
why this will feel familiar to many he said claude tag has completely changed the way i do work for the
last four months except it's not claude tag anthropic only announced that a few hours ago and i don't even
have access to it but i did build a version of it for myself which i've been using for months now
basically he said that after openclaw came online he built a harness that allowed him to turn any mac
into an ai employee with the claude code headless mode he then ended up building three such employees
and had them working in slack for the last few months now this idea of agents inside slack is also
something that we've experimented a lot with over here as well one of the reasons that i burned through
a billion tokens in march is that i was experimenting with a series of ai enablement agents that lived in
slack and could do things like figuring out how we were using ai providing coaching on how to better
use ai and developing overall ai strategy now you also might remember me talking about a month ago
about how the team at every had shifted their agent strategy from originally having every employee build
an ai agent version of themselves to instead having a group of agents that worked across employees based
on function and goal and had shared context claude tag in many ways seems to be institutionalizing some
of those shifts as best practice so trying to sum this all up let's talk about the five big shifts
that i think claude tag could represent and keep in mind this is not just about claude tag as a feature
but about it as a leading indicator of how labs and agent app companies are thinking about the future of
ai at work the first big shift is from an app native interface to existing workplace interfaces obviously
this breaks agents out of the controlled settings of the web app or the desktop apps that the labs have
built and instead uses the existing spaces that people do work in now to some extent one has to
think that this was inevitable as you try to shift behavior and get everyone working in a new way
one of the biggest points of friction is the new interfaces that they have to go interact with
the fact that the full power of claude code and everything that it can build is now accessible from
the most default channel where people are working seems likely to me to make a big difference in terms
of who accesses that power likewise this moves ai from a private chatbot to a shared teammate experience
now of course there have been no shortage of sharing based features to break your personal
chats out of their home in fact both openai and anthropic have released versions of better native
websites slash web app builders in order to help make sharing more easily this of course takes that
to the next level which also leads to our third shift which is from single user context i.e what the
ai or agent knows about you to the full team context part of the power of moving the agent from its native
home to your team's native home is that it gets access to the ambient context that exists all around
it without you having to create an entirely different way of accessing all that context now already
claude tag also reinforces a shift which was already very much happening which is the shift from
prompting to delegation as the main mode of interacting with ai already throughout the year as agent
the capacity has come online people have been slowly but surely breaking out of the way that they used to
interact with these tools a big part of that has been in shifting from telling the ai what to do to
telling it what you are trying to accomplish and giving it more freedom and latitude to accomplish an
ever-expanding scale of goals lastly claude tag and more broadly agents that work in your team's workspaces
shift ai's importance from something that becomes personally essential to key workers to an actual
organizational dependency a new way of working not just individually but across the entire team
which is not to say that there aren't challenges first of all while yes this is theoretically just
dropping claude into slack there are a lot of questions around access and tool configuration that
make it set up potentially much more complicated than it at first seems anthropic released an entire
blog post about best practices for setting up agent identity in this new claude tag paradigm
they also pointed out some best practices around permissions which wouldn't necessarily be obvious at first
glance gail wiener points out that while the positive side is the democratization of claude
code capabilities to a broader subset of the organization the flip side of that is that something
like claude tag potentially runs into existing trust barriers and skepticism that could create a challenge
she describes a scenario where there are five people on a team with her as the ai power user two
members of the team not trusting ai one team member thinking that the power user uses ai too much
and another being indifferent gail writes i start tagging claude in the project chat and here's what
happens in that room the moment claude is in the channel every message anyone types is being read by
it it stops looking like a tool the team uses and starts looking like my tool that's now sitting in
everyone's workspace i become the person who brought the surveillance device to the meeting even if that's
not what it is the skeptics position gets validated every time claude produces something good see she's
outsourcing her thinking and every time it produces something mediocre see this is what we were worried
about you can't win that frame basically she's arguing that this creates a whole new challenge around the
human layer that will have to be solved as well one specific thing that seems fairly challenging to me
is that despite the fact that you just use at claude for every instance of where you want to call upon
these capabilities different clauds and different channels have potentially access to different tools
and different permissions and different contexts and that feels like it'll get confusing super super
fast in fact in simon smith's first impressions post later on he did write claude being many clauds is a bit
disorienting i've come to think of claude as my claude the one in the app that knows me and connects to
things i've connected it to but there is no single claude in slack it's channel dependent and admin
configured and none of the clauds are mine this seems to be a common feeling one of the first things i'm
seeing is people tagging claude in channels to ask what it can do what it remembers how it maintains
context in the channel indeed how this creates new challenges and opportunities for context is a big topic of
conversation and for some in light of the fable challenge in light of everything that's happened with
fable there's also increasingly good reasons to not want to outsource this sort of process to a single
closed provider hugging faces head of product victor writes at hugging face we've been building our own
agent that we use via slack honestly building your own is quite simple and you'll be happy you did
any model you want including self-hosted if needed fully customizable to your stack and of course no
lock-in and no waitlist and not overpriced look ultimately it is very easy to overestimate the
significance of any one new feature and yet if for no other reason than these significant
shifts that the anthropic team themselves are reporting it seems like claude tag is one paradigm
shift that will at least be worth exploring i'll report back as we try it out and see what we learn
for now though that's going to do it for today's ai daily brief appreciate you listening or watching
as always and until next time peace