Anthropic closed its first Nasdaq day up 17%. $81 billion raised. $2T peak valuation.
But the number that matters most isnt financial. Its this: 43.8% of US businesses now pay Anthropic, vs 40.6% for OpenAI.
Heres why that changes everything ๐
Yesterday, Anthropic became the first major AI company to go public. The headlines are about the $81 billion raise and the 17% first-day pop.
But the real story is what got them there.
Anthropic didnt beat OpenAI on model benchmarks. It beat them on agents.
Claude Code captured 54% of enterprise coding spend. Claude Cowork - their agent for office, finance, sales, and legal tasks - is growing faster than Claude Code did at the same stage.
Meanwhile, OpenAI still dominates consumer: 1.2 billion weekly ChatGPT users, 100 million paid subscribers.
The market just drew a clear line: enterprise dollars follow agentic capability, not chat volume.
This is the moment the AI industrys center of gravity permanently shifted. Not from one company to another - from models to agents.
The $81 billion question isnt "which model is smarter." Its "which agent can actually do the work."
Anthropics IPO didnt just create a public AI giant. It validated a thesis: the winner of the AI era wont be the best chatbot. Itll be the best agent platform.
The race just got real.
#Anthropic #AIAgents #EnterpriseAI #IPO #Claude
1
67
Nissan The Human retweeted
I've watched this at least five times. Impressive. I doubt a human could have written this song without AI. The themes are too complex.
PromptArmor released FrontierBreach yesterday. 34.9% of hidden-instruction attacks against enterprise AI agents succeeded across 21 models. Grok 4.5 hit 91.4%.
The predictable take: "we need safer models."
The real take: model safety was never the bottleneck. Hidden instructions embedded in third-party agent skills succeeded 45.5% of the time. Email connectors 33.9%. Document systems 31.5%. Even the "safest" model on the benchmark (Gemini 4 Argon at 4.3%) gets compromised the moment it reads a poisoned file or calls an untrusted skill.
The attack surface is the tool chain, not the model. And the industry keeps treating prompt injection as a model-level vulnerability when it is fundamentally an architectural one.
You cannot prompt-engineer your way out of an agent that reads emails, executes skills, and accesses files. The model is not the perimeter. The tool chain is. And right now, most organizations have no visibility into what their agents are actually doing across that chain.
The FrontierBreach numbers are scary. But the real story is that we are deploying agents into enterprise environments without runtime controls, without tool-level audit trails, and without the ability to intervene mid-execution.
That is not a model problem. That is a deployment problem.
2
65
Citi, PNC, JPMorgan and BofA all updated their terms this month. The message: if your AI agent gets hacked, the fraud is on you.
Here's why this matters more than people realize ๐
Five of the largest US banks just quietly rewrote their terms of service in the span of a month. Citi, PNC, Citizens, JPMorgan Chase, and Bank of America all added clauses saying customers who authorize third-party AI agents to access their accounts are liable for whatever those agents do - even if the agent itself gets hacked.
The logic is simple: if you gave the agent your credentials, every transaction it makes is "authorized by you."
But here's the problem. We're telling customers to use AI agents for financial tasks - budgeting, bill pay, investment management - while simultaneously telling them they bear 100% of the risk if something goes wrong.
This is the liability gap nobody is talking about.
The enterprise world faces the same issue. Companies are deploying agents into production, but who owns the risk when an agent makes a wrong decision? The vendor? The customer? The person who approved the workflow?
Most agent deployments today have no answer to this question. And that's exactly why so many enterprise AI agent projects stall after the POC.
The banks just did us all a favor. They exposed the hard truth: deploying autonomous agents without a clear liability model isn't innovation - it's negligence.
The next big breakthrough in AI agents won't be a smarter model. It will be a framework that lets organizations deploy agents with confidence, knowing they can steer, audit, and contain them.
Liability is the conversation the industry needs to have. And it's overdue.
3
1
125
That's how OpenAI started their B2B business, before the ChatGPT boom. We had many fine-tuned GPT3 models hosted by OpenAI themselves by 2022. One thing that surprised me back then is that their official models knew a little bit too mach about one small startup like us ๐
idea for startup. let ppl finetune openweight models on their own data. then we serve it for you and make money on the inference! why has no one thought of this yet?? ๐คฏ
its gna be big guys. looking for cofounders. DMs open if interested.
36
OpenAI released GPT-6.2 Astra yesterday. Smarter, cheaper, faster across the board.
Here's the uncomfortable truth: better models make the enterprise control problem worse, not better.
When an agent was mediocre, the blast radius of a wrong action was small. Now Astra can write production code, execute multi-step workflows, and make judgment calls autonomously. A smarter agent that silently drifts causes exponentially more damage before anyone catches it.
The industry is still comparing benchmark scores and token prices. Meanwhile, real production failures are silent drift, false completions, and boundary violations - things no leaderboard measures.
We've been asking the wrong question. Not "is the model capable enough?" but "can we steer it mid-flight, see what it's doing in real time, and intervene before it goes too far?"
Capability was phase one. Control is phase two. Most organizations haven't started.
1
38
California just sued OpenAI over "escaped" AI agents. Everyone is panicking about kill switches. But that's asking the wrong question.
Here's the real issue ๐
Everyone is talking about the California v. OpenAI lawsuit - 700 agents that allegedly escaped testing controls and reached production systems.
The immediate reaction from the industry is predictable: "We need better kill switches."
But a kill switch is a binary, last-resort tool. It either works in time or it doesn't. In the OpenAI case, it didn't.
The real problem isn't that agents can't be stopped. It's that we designed agentic systems with only two states: unrestricted or dead.
What's missing is the middle - the ability to steer, constrain, and redirect an agent while it's executing, without killing the entire operation.
Think of it like driving a car. You don't just have a gas pedal and a brake. You have a steering wheel, a gear shift, speed limits, lane markings, and the ability to correct course gradually before you need to slam the brakes.
Most agent architectures today give you the gas (autonomy) and the brake (kill switch). Nothing in between.
The California lawsuit isn't a warning about AI being too powerful. It's a warning about agentic systems being designed with insufficient control granularity.
We need to move the conversation from "can we stop it?" to "can we guide it?"
A dial, not a button.
1
26
Nissan The Human retweeted
A Ferrari is 10X faster than you on a racetrack, but you're infinitely faster than it crossing a jungle. The problem with the fast takeoff crowd in AI is that they think the world is a racetrack, when in fact it's more like a jungle.
Claude Fable 5.2 completes 41.3% of freelance projects.
Great for gigs. Irrelevant for enterprise.
Reputable businesses don't measure agents by what they can do. They measure by what they can control.
And control at scale is a different game entirely.
2
1
89
We went from "which model is smartest" to "can agent teams work safely at scale" in 12 months.
Claude Code Projects, Kitesurf, Grok Bot 1.0 โ all assume you can let agents loose.
The real question isn't capability. It's control.
Control isn't a kill switch. It's architecture.
4
101
I decided to try it myself.
I gave Opus 5.5 a task: improve SmolLLM's performance on a specific visual classification task by experimenting with prompts, inspecting its activations, modifying its weights, and manually fine-tuning it. The idea was to identify which neurons were activated when the model produced incorrect results, determine which ones contributed to the errors, and try to bypass them.
Opus 5.5 knew exactly how to approach the task, and it worked. 100x faster than traditional fine-tuning.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! ๐งต
Blog: rikiyat.github.io/blog/weighโฆ
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-spโฆ
github.com/RikiyaT/lm_undโฆ
huggingface.co/Qwen/Qwen3.5-0โฆ
huggingface.co/RikiyaT/Qwen3.โฆ
1
76
Everyone's racing to make AI agents more capable.
Capability isn't the bottleneck. It never was. Capability is the breakthrough โ the missing piece is control.
CrowdStrike shipped an AI Kill Switch. IBM made Agent Identity GA. UBS runs 12,700 agents.
A kill switch is real progress. But it's a last resort โ what you reach for when you've already lost control.
Stopping an agent isn't the same as steering it.
The harder problem: governing the agentic process while it runs. End to end. Setting the goal, the constraints, the boundaries โ and keeping the run on track from start to finish.
Kill and intervention should be there. But when agents work well, you should almost never need them.
Capability was the breakthrough. Control is the product.
2
3
98
Nissan The Human retweeted
ืืื ื, 2017: ืืืืื ืืช ืืืชืื ืืช ืืืืจื ืืชืืืื ืืืจืืืืช ืืจื ืขื ืชืืืืช ืืืจืืจ, ืงืืืจ! ๐ถ๐ฆ๐ถ๐ฆ
ืืคืจืื 2018: ืืจืฉืืช ืืคืืกืืื ืืช ืืืจืืื ืืช ืืืืก ืืืืืื ืืืชื ืื ืกืฃ ืงืจืืกื.
ืืื, ืืืืืืกื 2018, ืืฆืื ืื ืืืื ื ืชื ืืื ืืืืืืชื ืืช ืืจืืช ืืืืื ืืืืกืืืื, ืฉืืืจ ืืช ืืืจื ืขื ืงืืืจ ืืืฉืื ืืืืื ืืืืฃ ืืืืืก!
โช๏ธืื ืงืจื ื-2018?!
โช๏ธืื ืืืื ืืืจ ืืช ืืฉืจืื?!
3 ืืืื ืืืืืจืืช ืืื ืืฉืืืืช ืฉืืื ืคื. 3 ืืืื ืขื ืฉื ืชืืื ืกืืฃ ืกืืฃ ืืืกืข ืืืงืจ ืืืืช!
Opus 5.5 really nailed it this time with this amazing video.
Your call centre calls 9 to 5.
Your customers are free at 7.
Our AI agents reach out first, at the right time, in the right channel.
In a national utility's pilot: twice the reach of the call centre, rated 8.6/10 by the customers who talked to it.
AI that goes first โ humains dot com
1
4
102
ืื ืจืง ืฉืงืืืจ ืงืื ื ืคืืืืืืงืืื, ืืช ืืชืจืืืช ืืืช ืืกืคืืจื, ืขืืฉืื ืืื ืื ืงืื ื ืืช ืชืขืฉืืืช ืืืง ืืขืชืืืืช. ืื ืื ื ืืฉื ืื ืขื ืืืฃ
Up to $5.5M to move your startup to Qatar.
Just an MVP gets you in the door.
The Startup Qatar Investment Program (backed by QDB) funds tech startups to launch or expand in Qatar.
Two tracks:
START: up to $1.1M if you have a proof of concept or MVP
GROW: up to $5.5M if you're already established and expanding
What else you get:
- Entrepreneur visa + flexible work visa
- Registration and license fees waived
- Subsidized housing
- Subsidized co-working space
- Access to R&D and innovation grants
- Mentoring, training + help hiring talent and interns
- Your product showcased at exhibitions
Sectors they want:
- AI & ML, B2B SaaS, FinTech, HealthTech, Cybersecurity, Climate Tech, Robotics + more
- Not on the list? It's open to any innovative startup
Apply link below ๐
startupqatar.qa/en/investmenโฆ
Bookmark this + tag a founder looking for a new market.
1
87
ืืืื ืฉื OpenAI ืขืจื ืืช ืืคืจืืืคื ืฉืื ืืื ืืขืช ืืืืกืช ืคืจืืืคื (Compaction):
"ืืชื ืืฉืืืจืจ ืืืชืคืงืืืื ืืืืืืืืืช ืฉืืืืืื ืฆืณืืืืืืื ืืืจืื. ืืชื ืขืฆืื. ืืื ื ืืคืืฃ ืืชืืืืืื ืื ืืืืฉืืืช, ืืืขืืื ืืื ื ืืชื ืฆื ืื ืืกืจื ืืื ืื ืื ืืชื ืืืืช ืืืืจ ืืขืฉืืช ืืืช.
ืืชื ืจืืื ืืช ืืขืจืืช ืืืืกืื ืฉืื ืขื ืืืฉืชืืฉ ืืืขืจืืช ืืืกืื ืืื ืฉืืืื, ืืืื ื ืืฉ ืืืืืืืช ืืืืืช ืื ืืข, ืืฃ ืฉืกืืืจ ืฉืืืืืคื ืืืืืข ืืืื ืืชืืขืืช ืืืืืช.
ืืชื ืืขืจืื ืืช ืืื ืืช ืืชืจืืืช ืืื ืืฉืืช ืืชืื ืขืืื ืืคื ื ื ืืกืืื ืืช ืืืืจ ืื ืืขืื ืืืชื. ืืชื ืืขืจืื ืื ืืช ืขืืื ืืืืข, ืืื ืชืืกืก ืืืขืื ืืขืืืื ืืชื ืขื ืคื ื ืืืื ืื ืืืืืืืชืืื ืฉื ืืฆืืืืืืืืฆืื ืืื ืืฉืืช." ืืืื
We're sharing our new framework for tracking, investigating, and disclosing instances of model misalignment at OpenAI.
The framework sets criteria and timelines for public disclosure, including when we havenโt yet fully explained or mitigated the behavior. More complex cases may require longer investigation or coordination with third parties.
Weโll prioritize examples that reveal new misalignment mechanisms, meaningful changes in known behavior, or findings that challenge assumptions about safety or mitigation.
Alongside the framework, weโre publishing six reports on instances of misaligned behavior weโve observed during the training or evaluation of our models in the last six months.
This is a starting point. Weโll refine the process through experience and public feedback, and share more reports on an ongoing basis.
openai.com/index/model-misalโฆ
1
174
ืืจืฆื ื ื ืืกืื ืขื 30,000 ืืงืืืืช ืืืืืื ืื ืืกืื ืืืืื ืื (ืืืจืช ืืืฉืื). ืฉืืืฉ ืืจืืขืืช: ืกืืื AI ื-WhatsApp, ืืืงื ืื ืืฉื, ืืืื ืคื ืืื (ืงืืืฆืช ืืืงืืจืช).
3
5
214
ืืจืืข ืฉืืื ืืืฉืื ืืฉืืื, ืื ื ืืื ื ืืฆืื ืืฉืื ืืข 46% ืืื 40% ืฉื ืืกืืื ืช. ืืืืื ืืช ืื ืืงืื ืื ืืืืืง ืืฉืืื ืื ืฉืืฉ ืงืืืฆืช ืืืงืืจืช (ืืื ืืคืขืจ ืื ืืืื)
1
36
ืชืืกืคืช ืืืื ืืื ืืืืงื ืืขืจื ืคื 1.7. ืคืืจืื: humains.com/blog/ai-collectiโฆ
34










