Tim Hua 🇺🇦

I gave @GiveWell ten thousand dollars and all I got was this hat

I dislike Smith implicitly encouraging OpenAI to communicate less honestly in order to be more successful. Second post this week.

They did, it was called America

as more excellent poasters get sucked into research we will see papers titles with mimetic strength previously assumed impossible to the lay academic

@MattZeitlin Coke Zero drinkers:

We’re calling it: The war in Ukraine has entered a new phase. @KatStepanenko and I have authored a new special report studying how Ukraine is actively challenging the positional character of the war that has dominated the battlefield since 2023. Data on Russia’s battlefield performance indicates that the character of the war is shifting in favor of Ukrainian forces – at least for now. Russian forces rates of advances are stagnating while Ukrainian forces are employing novel tactics and operational concepts in efforts to break out of positional warfare. Neither Russia nor Ukraine can conduct operational maneuvers yet, however. The bottom line is that the war in Ukraine is competitive and far from stalemated. Ukrainian forces are out-innovating Russian forces in both military technologies and in applying these new technologies in effective operational concepts that can help Ukrainian forces break out of positional warfare. Ukraine is employing mechanized equipment in tactical maneuvers in ways that were impossible 12 months ago. Russia’s ability to conduct infiltration missions will likely continue to degrade as Ukraine’s intermediate-range strike campaign pushes Russia’s logistics and forward operating bases further away from the frontlines, reducing resourcing to sustain infantry tasked with infiltration missions. Ukraine may be able to scale these effects if they resourced properly by international partners. Ukraine’s advantage in intermediate range strikes is notably not permanent, and Russia will very likely eventually develop countermeasures to mitigate Ukraine’s advantages. Ukraine’s international partners thus have a rare and temporary opportunity to help Ukraine exploit favorable battlefield dynamics while Ukraine has the upper hand. Key Points of the report: • Russia’s rate of advance is plummeting during the Russian Spring-Summer 2026 offensive. • Russia is losing more soldiers to make fewer gains, with monthly Russian casualty rates reportedly outpacing monthly recruitment since December 2025. • Ukraine is starting to regain more ground than it is losing for the first time since 2023. • Ukraine’ recent counterattacks feature unique characteristics and deviate from key trends that defined the positional character of the war since 2023. • Ukraine is conducting a pattern of more frequent mechanized counterattacks at the tactical level for the first time since 2023. • The Ukrainian command’s operational planning is maturing. • Ukraine’s early 2026 counterattacks in the south were successful likely due to better planning and preparation of the battlefield. • Ukraine has been conducting a coherent campaign to suppress and destroy Russian air defenses since late 2025, in order to shape the battlefield as part of more sophisticated campaign planning. • Ukraine significantly intensified its intermediate-range strike campaign against dynamic targets in Spring 2026 in order to degrade Russian logistics at operational depths ahead of planned Ukrainian maneuver. • Ukrainian forces started actively disrupting Russian railway logistics in occupied Ukraine and Russian western regions in Spring 2026. • Ukrainian intermediate-range strikes are already achieving notable operational effects, including degrading Russia's ability to use the key Russian highway connecting Russia to occupied Crimea and GLOCs around Donetsk City. • Ukrainian forces decisively seized the initiative in intermediate-range strikes by fielding new technologies such as the US-made Hornet strike UAV, among other systems. • Ukrainian forces are achieving temporary tactical drone overmatch in some frontline sectors, which is slowing Russian offensive operations by degrading the effectiveness of Russian shaping operations. • Ukrainian forces likely achieved tactical drone overmatch in certain frontline sectors after degrading Russia’s drone capabilities in late 2025 to early 2026 - primarily by suppressing drone launch positions and increasingly intercepting Russian tactical UAVs. • Ukraine’s degradation of Russian forces at operational depth combined with tactical-level drone overmatch likely is creating vulnerabilities in the Russian lines. • Ukraine’s intermediate-range strike campaign is likely far from its zenith, assuming continued support from Ukraine’s partners, and will likely intensify over 2026 as Ukraine fields new weapons capable of striking Russian’s operational rear. Link to full report: understandingwar.org/research/russi…

@SimonLermenAI Ah yeah "it's also crazy" is a bit vague. Will try make it more clear next time

@Tim_Hua_ Took me a while to understand this tweet is about plot formatting (the lower row) Interesting that verbalized awareness is sometime higher than their probe thing.

This graph from the NLA paper, imo, provides pretty convincing evidence that activation verbalizers surfaces unverbalized eval awareness. It is also crazy. Notice how the verbalized eval awareness dot is offset only when it's significantly more than zero. How did they even...

made a site that picks the closest rothko for how the weather feels outside your window

I guess I shouldn't be too surprised given that there's a town called Dildo nearby

Bro you wouldn't believe the name of where we found the iceberg.

@a_karvonen hopefully one day! I wanted to get away from SF for this trip so I ruled out the entire west coast...

@Tim_Hua_ Check out Prince William sound in Valdez Alaska, I did a similar tour which was very scenic, in a narrow bay with steep mountains

I am currently vacationing in the center of the universe (St. Johns, NL) 🧵

... plus a dot.

...make a graph like this?? The off-set confidence intervals on the bars are crazy. On the one hand I feel like I can't complain because I did understand the graph, but on the other hand, I think they really should've just did like four bars per condition instead of two bars...

Maybe I'll try build some high quality evaluations for reward seeking/grader sycophancy behaviors. I bet that'll be pretty sick.

...doing what gets rewarded when that contracts the other stuff! I feel like models don't actively training-game coherently to the extent that it doesn't do anything super coherently.

Opus 4.6 would output even numbers in the extremely-toy reward-seeking evaluations from Nitishinskaya & Schoen, and this behavior goes up if you increase the amount of reward it gets. (From the NLA paper.) I've always felt like Anthropic models are very reward-seek-y...