Michael Teper (@mteper) profile picture

Working at Google, opinions my own.

Joined February 2009
Photo by Michael Teper (@mteper)Photo by Michael Teper (@mteper)Photo by Michael Teper (@mteper)Photo by Michael Teper (@mteper)Photo by Michael Teper (@mteper)Photo by Michael Teper (@mteper)Photo by Michael Teper (@mteper)Photo by Michael Teper (@mteper)Photo by Michael Teper (@mteper)Photo by Michael Teper (@mteper)
Michael Teper retweeted
everyone coding with agents should internalize this - the RLVR in current frontier LLMs' post-training is done at task level. in other words, they won't give a damn about long term maintainability this is why we see so many useless comments and test cases being written by default. these things make future tasks less efficient, but they don't affect the current task's success, so they survived the training and became part of the models' behaviors until post-training can reward long term maintainability (which i suspect is very hard because inherently it's a slow and expensive process to measure that), it's on us humans to correct it in the harness layer through prompts, skills, lint rules etc this also implies that understanding long term implications of technical choices is a valuable human skill that's hard for an average person to make up for with LLM it's the distinction of whether you can create a durable growing codebase, or have one that gradually degrades into insanity if you have that skill, you are golden (for now, until post-training catches up)
65
36
10
610
22,540
Michael Teper retweeted
I'm not Jewish. But one thing I figured out a long time ago is that any society that treats its Jews badly is about to become an unlivable shithole for anybody with an IQ above 120. Good enough reason to support them.
I hope my non-Jewish friends know how much I love and appreciate them for their unwavering support.
572
616
86
5,985
130,660
Replying to @mihaimaruseac
here you are
56
15
15
523
93,238
Michael Teper retweeted
Responding in my personal capacity, but wanted to share my perspective.. We are in a great-power competition with so much at stake in the years ahead. There is no one in America who has lived during a time when the United States wasn't the preeminent superpower and, despite our imperfections, a force for good in the world. A nation that fought and helped end wars on behalf of those who couldn't, pioneered powered flight, put humanity on the Moon, led revolutions in medicine and technology that improved billions of lives and distributed hundreds of billions in aid around the world. I don't know how handicapping this nation by going after one of its largest employers, a company that has created wealth for millions, ushered in electric cars, rockets that don't need to be thrown away, satellites that connect the disconnected and have unquestionably saved lives, not to mention access to information that is pretty foundational to solving many of the world's problems, biotech to restore function to those with disabilities, and so much more, serves America's interests. Without the contributions of these companies, there is little question America would be behind in a competition where the outcome is far from certain and the consequence so great. Now, maybe I am just a bit new to this, but from the billboards, TV and digital ads, spam texts, and fundraising calls, it would seem both parties have no issue raising money, including from billionaires, with the aim of influencing and buying votes. And if that means campaign reform, maybe include dialing back the spam calls and texts..But with all that is going on in the world, and so much at stake for our country, I question the priorities reflected in the post below.
We’re going to do everything we can to make Elon’s worst nightmare come true next Congress: a democracy where the ultra-wealthy can’t buy elections.
1,333
4,952
399
33,381
1,033,372
Michael Teper retweeted
When you hear hoofbeats, think horses, not zebras. The question of the past 5 days is whether near-simultaneous system outages across infra, transportation, communications, etc. are horses or zebras. Enough people on many sides of these issues (defense hawks, anti-AI, doomers, anti-tech, IT people, etc.) can find their own common threads across MSFT, FAA, United, Chase, NSA, AAPL and turn them into a theory. Depending on your perspective it explains things or seems like a (wild) conspiracy. If you're looking for horses, the common thread is approaching the classic system lockdown of pre-Thanksgiving through end of January. It would be very straight forward to explain all of these as just gearing up for that travel/retail/operational load and lockdown. The AI line through this is two-party. First, that many vendors (and OSS projects) have been using AI to find potential vulns and patch them. It is reasonable to assume the volume of patches has increased as the find rate has and the knowledge that once found in-house they were found outside. The volume of these fixes is likely stressing systems and so compatibility or breadth testing might not be a higher priority than getting fixes out there even if they break stuff. Second, the client/customer side of this should be (but might be) also using AI to test their infrastructure. They are likely finding many configuration holes in networking and permissioning. Changing these is always non-zero risk even in the best of times and trying to push those through before next week is a must do knowing again that adversaries are also looking and downtime during the holiday is a true disaster. In addition, the patches from vendors are being applied in essentially real time. There's no other choice as the patches coming out now lack the time and pacing to obscure what they are fixing making them information rich for bad actors. Neither of these preclude being closer to zebra hoofbeats—meaning adversarial attacks. The zebra scale can be anything from a single coordinated actor to literally a bunch of people all with the same tools pointed at different targets. The latter is possible because these tools all make it to all the curious/bad actors at the same time. That was the Friday FBI warning (ic3.gov/CSA/2026/261113.pdf). It is not much of a reach to think independent actors all started testing this out. Apple, Microsoft, Chase are obvious kinds of targets for stuff like this. But realize those are also exactly the entities moving quickly to get to a patched and locked down state. The most common thread is that all the systems are big and complex with the fragility primarily driving not by "bugs" in layers of the code but by configurations and permissioning that are extremely difficult to test and get right, and also very rarely reproducible by any vendor in the mix. This can broadly be called a challenge with legacy or just big systems. Overlay the new reality that AI is in fact driving which is an increased pace of identifying, fixing, and deploying code changes. We are collectively going to go through many more system downtimes than previously experienced until the technology stack (IT people and infra) get better at this new speed. This is exactly what the industry went through going through Y2K and just after with Windows XP. The past two decades, SaaS and mobile included moved at a well-rehearsed and comfortable pace. I discussed that here -> nitter.cf/stevesi/status/2101886… The FAA one stands out as not fitting this patten nearly as much. This one deserves more investigation because it involves new hardware. That's a known risk across all vendors but especially this type of hardware and even vendor(s). The diversity, timing, and apparent type of complex issue point to much more classic hoofbeats. It is easy to see the timing and assume coordination but the timing can also be because of availability of tools and the holiday. PS: CNN reporting that echo-sigma .com sales are +800%. As a prepper product of the cold-war it is fair to say that unless you've been preparing a backpack with a tube tent, multi-tool, and 3 days of life boat rations will not be of much help. Covid taught us that.
Enough debating what AI is or might become. It’s time to build on what it is today. An AI Action Plan: 10 concrete work items for Labs, enterprises, developers, and government—security, agents, access, liability, copyright, & more. Let’s get engineering.
6
32
7
144
22,520
Michael Teper retweeted
When the rocket blows up, it's "Elon Musk's SpaceX". When the rocket is fully successful, it's just "SpaceX". This is a demonstrated pattern across all major media and also within the NYT specifically.
63
869
34
8,556
77,126
Michael Teper retweeted
Will the future that you want
Choose your path Western man.
2,479
7,733
405
97,560
4,604,797
Michael Teper retweeted
Both the Biden and Trump administrations allowed the export of less powerful GPUs to China while reserving the most advanced chips for America. This has been a clear national security success. Per Grok and Claude, the US-to-China AI FLOPs ratio has grown from 2.8x in 2023 to 8.8x in 2026 and is modeled to reach 14.1x in 2027. Our advantage in compute has grown dramatically and American AI labs have stayed well ahead of Chinese labs the entire time. Chinese labs prefer these less powerful American GPUs to domestic accelerators from Huawei and others. That preference has slowed China’s hardware industry: AI chips improve through intensive use by leading labs, and the ready availability of American alternatives has limited that feedback loop for Chinese AI hardware. Chinese labs can also access capacity in Southeast Asian clouds run by American hyperscalers. Those zones can be shut off instantly in a conflict, giving the US signficant leverage. This dependence is advantageous in every way. A complete ban on GPU exports and Southeast Asian cloud access might be deeply damaging to long-term American national security. It would force Chinese labs onto the Huawei AI hardware ecosystem. China is not power-constrained, and Huawei is developing extremely power-hungry, large optical scale-up domains that may approach American state-of-the-art systems through brute force rather than architectural elegance and efficiency. And China has a lot of power. Having Chinese labs optimize models for this different architecture is risky as nothing prevents China from sharing this hardware and associated models with Russia, North Korea, or Iran. We would lose any ability to cut off their AI in a conflict. Policies now under discussion in Washington carry serious, uncertain risks. Reasonable minds can disagree, but treating this as a simple black-and-white issue does America a disservice. If China closes the installed FLOPs gap as its domestic hardware matures rapidly as a result of increased Chinese lab usage after stricter export controls, we will know this new policy was a mistake. If Russia, Iran, or North Korea later field comparable Chinese AI hardware systems that are used to run AI targeting models against Ukrainian or American troops, we will know stricter controls were a *terrible* mistake. The current reality is one where China is consistently behind America thanks to the export of less powerful GPUs and we have immense leverage over China in any future conflict. We know that the current policy is working. We have no idea what the consequences of stricter controls might be. This is why China is currently restricting imports of roughly 600,000 H200s sitting in Hong Kong warehouses. China Hawks should consider why Beijing agrees with them and why China *restricts* the sales American GPUs inside China. The current policy has worked spectacularly. The alternative is extremely risky with unknown consequences.
143
231
53
1,647
335,252
Michael Teper retweeted
Mechazilla catches Starship! Congratulations to the entire SpaceX team on this incredible achievement!
6,798
27,379
9,114
141,312
17,649,320
Michael Teper retweeted
Oh, and in case you're wondering - there's much more to unpack. Much, much more! I just haven't had the time for all of this! The gravity and sensitive nature of this kind of data is beyond imaginable. (I reckon that some of this can give analyst a PTSD ;) )
The European Commission is ordering Google to share the search data of all Europeans with other companies (AI chatbot makers, search engines, etc.), after a kind of "anonymisation". I tested it on the real Google data, hundreds of billions of search records. I broke it. Findings. I singled out the people behind half a million searches, and that only for few countries over one month of August (record isolation). I joined the pieces of search history beyond the defined cap of 3, more than 200,000 sets just for Germany and only in August (linkage). I also confirm that it is possible to perform inference on the dataset, for example concerning health, religion or political opinions. This moves us beyond data protection, and in the national security direction. My report: "Alone among a thousand". (link below)
4
23
2,061
Michael Teper retweeted
Big G is back.
What we shipped this week @google Monday: Two new 8th-gen TPUs Tuesday: Gemini 4 Carbon - AlphaProof Code verified Signal protocol - Gemini Robotics 2.5 Wednesday: AlphaGenome 2 for disease research - Gemini Decisions API in AI Studio Thursday: Gemini 4 Hydrogen - intelligent eyewear pre-orders Friday: U2 new album (ok not us but exciting!) onwards!
7
5
3
237
17,785
Trump told JD Vance to focus on domestic issues after reports that Vance was considering a meeting with former IRGC chief Mohsen Rezaei. Iranian historian Shahram Kholdi explains what Rezaei’s emergence tells us about who holds power in Tehran. Freedom advocate Gazelle Sharmahd asks why Washington should negotiate with the Islamic Republic in the first place. Whoever gives the orders, she argues, the regime poses a threat to American security and “must be brought down.” You can watch the full episode on YouTube or listen on any podcast platform. @MENAResearcher @GazelleSharmahd
8
25
3
99
13,108
Michael Teper retweeted
You can long-press to make a text selection but there is literally nothing you can do with the text
44
35
10
1,732
82,698
Michael Teper retweeted
I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo…. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
1,822
6,424
2,209
41,251
6,787,286
Michael Teper retweeted
Replying to @ATLCWorker
The poor should be afraid of socialism because it guarantees that their poverty will never end.
15
65
3
1,527
17,513
Michael Teper retweeted
This is the definition of insanity, and it shows how far the @UN has fallen.
SHAME: The U.N. General Assembly just elected the murderous Islamic Republic of Iran to its Committee for Program and Coordination, the main body shaping the world organization’s priorities on Women’s Rights, Human Rights, Disarmament and Counter-Terrorism. 144 voted for Iran. 🧵
1,110
5,306
187
27,207
462,502
Michael Teper retweeted
From the way everyone is reacting, you’d think Jensen strangled his three babies to death and then blamed it on postpartum psychosis.
263
5,031
132
64,483
1,222,673
Michael Teper retweeted
Lewis Hamilton, the Formula 1 driver and Kim Kardashian's boyfriend, visited the Egyptian side of the Rafah crossing this week, shown around by the governor of North Sinai. He wrote that he saw walls stretching for miles, built to keep Gazans in, and in the post to his 44M followers he blamed Israel. Except those walls are Egyptian. Egypt razed the homes of their own people near the border to make room for them, and reinforced them during the war so that Palestinians fleeing the fighting could not cross into Sinai. He stood in the country that built the wall, next to the official who runs the province it sits in, and never thought to ask who put up the wall or why. He's not wrong about aid and how Palestinians need it, but advocacy sort of loses its efficiency when you show that you are either entirely ignorant about the situation or lying to feed a narrative. Oh and Shawn Mendes reposted it too. I guess fact-checking is for a bygone era.
298
1,100
103
6,501
379,667
Michael Teper retweeted
What's the difference between masked bike riders targeting Jewish businesses and the KKK marching through a neighborhood to intimidate people they hate? The bicycles?
142
389
18
2,940
34,153