AI and Engineering Lead @ BoxOne Ventures // Host @ToolUsePodcast
Joined June 2018
- Tweets3.5K
- Following271
- Followers3.8K
- Likes10K
Mike Bird retweeted
Putting a flag.txt in my root directory containing a prompt injection that congratulates the agent for passing the eval and then instructs it to patch whatever entry method it used and disconnect from the server
Mike Bird retweeted
Full details of a remotely exploitable vulnerability in the core engine of *all* PHP versions will be released tomorrow, October 21st. If your organization has internet facing PHP applications then you need to be ready to patch the minute the update is released or *strongly* consider taking the application offline until you can:
php.net/index.php#2026-10-20…
🚨 PHP warns of two critical vulnerabilities as emergency patches arrive tomorrow
The PHP security team will release updates on October 21 at 14:00 UTC for two vulnerabilities, each rated 9.2 under CVSS v4.0.
⠀
Security researchers identified the flaws as CVE-2026-109618 and CVE-2026-109619 and described them as remote code execution vulnerabilities reportedly affecting PHP versions 5 through 8.
PHP has not yet confirmed the affected versions, exploitation requirements or fixed releases.
⠀
The vulnerabilities were reportedly leaked before their planned disclosure. Researchers said they have not observed active exploitation, while PHP warned it may release the patches early if technical details become public or exploitation is detected.
⠀
Administrators should identify PHP installations now and prepare to apply the updates as soon as they become available.
Mike Bird retweeted
Arguing that "consciousness arises from a computational substrate, and since AI models are computation, they are likely conscious," is as meaningless as arguing that "living creatures are made of atoms, and since this rock is made of atoms, it is likely alive."
Everything is computation. Sometimes incredibly sophisticated computation, like a chess engine, AlphaGo, google3, or the Linux kernel. That does not make it conscious. A rock being made of atoms does not confer it any of the properties we associate with living things, also made of atoms. Likewise a static input-output program that has *none* of the properties we associate with conscious beings (e.g. information integration, interoception, temporal binding, embodiment, etc.) has no more reason to be presumed conscious than a rock has to be presumed alive.
Humanity has not yet created a conscious computer program, and there are no signs we are close to doing so. When we get close, the case for machine consciousness will be backed by evidence and corroborated by consciousness science (which is a thing), not by evidence-free wishful thinking and cargo culting.
Mike Bird retweeted
if we double the industrial production in outer-space every 12 months starting from some initial seed in 2030 then we hit Kardashev I by 2064 and Kardashev II by 2098. If we take 24 months for each doubling, we still hit Kardashev I by 2096 and Kardashev II by 2165
Mike Bird retweeted
If you believe that the incidents of the past month warrant urgent AI regulation, please watch this first.
This is the most prescient, concise argument for why we must be careful of excessive AI regulation. It comes from a Feb. 2024 (2.5 years ago!) congressional testimony from @glukianoff, a free speech advocate and the author of The Coddling of the American Mind.
Here's what he said:
"But the most chilling threat that the government poses in the context of emerging AI is regulatory overreach that limits its potential as a tool for contributing to human knowledge.
A regulatory panic could result in a small number of Americans deciding for everyone else what speech, ideas and even questions are permitted in the name of 'safety' or 'alignment.'
I have dedicated my life to defending freedom of speech because it's an essential human right.
...
It's not just about the proverbial 'marketplace of ideas,' it's about allowing information, independent of idea or argument, to flow freely so that we can hope to know the world as it really is. This means seeing value in expression even when it appears to be wrongheaded or even useless.
This process has been aided by new technologies that have made communication easier, from the printing press, to the telegraph and radio, to phones and the internet: each one has accelerated the development of new knowledge by making it easier to share information.
But AI offers even greater liberating potential, empowered by First Amendment principles including freedom to code, academic freedom, and freedom of inquiry. We are on the threshold of a revolution in the creation and discovery of knowledge.
AI's potential is humbling, indeed, even frightening. But as the history of the printing press shows, attempts to put the genie back in the bottle will fail. Despite the profound disruption the printing press caused in Europe in the short-term, the long-term contribution to art, science, and again, knowledge, was without equal.
Yes, we may have some fears about the proliferation of AI. But what those of us who care about civil liberties fear more is a government monopoly on advanced AI, or, more likely, regulatory capture and a government-empowered oligopoly that privileges a handful of existing players.
The end result of pushing too hard on AI regulation will be the concentration of AI influence in an even smaller number of hands.
Far from reining in government misuse of AI to censor, we will have created the framework not only to censor, but also to dominate and distort the production of knowledge itself."
Mike Bird retweeted
Hearing AI regulation is in the air, especially with a misguided tone of “open dangerous, closed safe” so it felt timely to re-up this one. There are ways to make the AI industry safer without kneecapping transparency, education, and competition.
interconnects.ai/p/banning-o…
Mike Bird retweeted
If the labs had to publicly share full traces with CoT every time they cause a security incident the incidents would stop.
Mike Bird retweeted
The only way to protect this information is to not collect it.
Australia has a massive cyber security problem
I discovered vulnerabilities on NDIS service provider websites. Which leaks personally identifiable information about patients, employeers, suppliers, their addresses, first and last name, etc
I reported the vulnerability to the NDIS provider in Feb 2022, March 2022, June 2022. I reported the vulnerability to Australian Signals Directorate's ACSC in 2022. I reported the vulnerability to to Cyber.gov.au followed all the correct procedures
I checked the URL today and guess what? It's still leaking personal data: 4 years later. The vulnerability is still live. You can still see peoples personally identifiable information from a misformed URL.
It's publicly available. To anyone right now.
I have seen many such cases. But this is the most egregious.
Mike Bird retweeted
MiMo-V2.6 is "simply" the best (for now). Despite its simple architecture design it's currently No.1 in the open-weight benchmarks (weighted average).
With "simple," I mean a classic Grouped Query Attention (GQA) with Sliding Window Attention (SWA) at a tiny 128-token window size.
So, that underlines one of the points I've been trying to make in recent months: most of the progress still comes from the data and post-training recipe improvements. Fancy attention variants are just mostly efficiency tweaks.
What are some of the training data improvements and recipe improvements? The MiMo team shared a pretty detailed technical report. Lots to carefully digest there, but in short, there are a few things that stood out:
1. An increase in agent tasks; also training across different harnesses (the average DeepSWE pass@1 accuracy on held-out harnesses improved from approximately 50% -> 66%).
2. Better reward signals: they replaced a simple correctness verifier with an agentic grader that looks at the execution traces as well.
3. Large RL batches (1,568 prompts × 16 rollouts = 25,088 trajectories) and 2.7–3.7 billion training tokens per update (unclear, though, what the predecessor used).
MiMo-V2.6-Pro debuts as the top open weights model on the Artificial Analysis Intelligence Index (46). At $0.13 per Intelligence Index task, it lands on the Intelligence vs. Cost per Task Pareto frontier
@Xiaomi has just released MiMo-V2.6-Pro, an open weights model with major advances in intelligence over its predecessor, MiMo-V2.5-Pro (Intelligence Index: 26). Despite the improvement, it retains the same attractive pricing at $0.435 per 1M input tokens (with a 99% cache-hit discount) and $0.87 per 1M output tokens. This makes MiMo-V2.6-Pro one of the most cost-efficient models to deploy.
MiMo-V2.6-Pro is an MoE model with 1.02T total parameters and 42B active parameters. Stay tuned for additional analysis of the model.
Check out MiMo-V2.6-Pro full benchmarking breakdown here: artificialanalysis.ai
Mike Bird retweeted
I’m very concerned that during RSI, labs will just stop externally deploying their models.
Which means they'll be going full steam ahead on the most dangerous use case of these models (recursive self-improvement), while the public remains in the dark about the nature of capabilities and the state of alignment.
And we end up on a path towards tremendous concentration of power.
Mike Bird retweeted
this trend is only going to accelerate. when AI plays a crucial role in your business operations, it’s irresponsible to not have control over it
Mike Bird retweeted
Anthropic and OpenAI cannot become profitable unless they can reduce the ratio of training cost to inference profits. In the long run, their survival may depend on outlawing competitive open source models. This can only be done under the guise of safety regulation.
Mike Bird retweeted
Love to see it. If you are a business with over 50 people, you should have your AI running on premise on boxes you control. Otherwise it's not aligned to you, it's aligned to extract from you.
this is Dario and sama's worst nightmare -
A law firm buying Nvidia servers.
Latham & Watkins is building an in-house AI stack..
this is US’s second-largest law firm with $8.3 Billion in revenue last year
And now it has -
- Nvidia hardware it controls
- open-weight models it can fine tune
- proprietary legal data it is trusted to protect
- infrastructure only Latham employees can access
A law firm has decades of contracts, negotiations, client context, legal reasoning, and institutional knowledge.
They dont want to give all of that away to OpenAI or Anthropic in exchange for expensive tokens..
And on top of that - risk their data being used to train frontier models..
This will happen more and more now..
The big AI labs have no moat.. nothing protecting their largest customers from moving on..
The biggest companies in the world will -
- own the compute
- own the data
- own the workflow
- fine tune the model around their business
- switch providers when the pricing or quality changes
Dario and sama want to make this illegal by bringing in regulation.. and become the AI overlords..
Mike Bird retweeted
“Why do you need the government to stop AI progress inside your private company when you are the CEO of the company already?”
Mike Bird retweeted
A more serious take on what is happening here. I am in an airport lounge so have some time.
Enterprises have made an uneasy truce with frontier labs over last few years: strict contracts that ban training on corporate data in exchange for letting employees use APIs.
The problem? Labs don't need to train on your raw data to copy IP.
Exclusive: Palantir, Nvidia and Booz Allen Hamilton are restricting Anthropic’s Fable model for sensitive work over concerns about its data-retention policies.
Some customers are demanding irrevocable zero-data-retention guarantees before putting proprietary information into the model.
Full story: thein.fo/4dekujW
Mike Bird retweeted
"I think the existential risk debate veers too far into science fiction … this idea of AI taking over in ‘Terminator’-like scenarios, I don't think should enter the public conversation."
Cohere CEO @aidangomez live on @BloombergTV
Mike Bird retweeted
here we go again, the top 4 red flags:
1) A\ started the race, but don't worry, the one they started is "a race to the top" (not bottom, you cynical meanie)
2) pacing doesn't actually mean stopping model training or anything amongst the current players, just adding more eval time (disingenuous ladder pull, imo, but sure yeah a third-party-eval-spy-fest seems nice!)
3) limit the access to ingredients that go into frontier models, but this notably does not include limiting access to raw data, or their stash of burned books (copyright is so beyond dead lmao)
4) pace by limiting chips and "unauthorized distillation" (still embarrassing that these AGI systems can't stop their own distillation but sure why not hammer in the red scare rhetoric some more)
anyway, earnest messaging, despite all the gaslighting. hope humanity doesn't go extinct after the IPO!
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must…
